docs: refresh stale roadmap entries

[andygrove]

Which issue does this PR close?

No issue. This is a documentation refresh of the contributor roadmap, following the last pass in #5064.

Rationale for this change

Several roadmap sections still describe work as unstarted or blocked when it has since landed, and a few references point at issues that have closed or code that has moved. Contributors use the roadmap to find where work is coordinated, so stale entries send them to the wrong place.

What changes are included in this PR?

Only docs/source/contributor-guide/roadmap.md changes. Each section was checked against main and the linked issues and PRs:

• Iceberg Table Writes: the section said no design was committed. The split writer/committer plan (feat: Optionally split the Iceberg V2 write operator into distinct writer and committer operations #4658) and the iceberg-rust writer (feat: implement native Iceberg V2 writer via iceberg-rust #5361) have landed behind two experimental, off-by-default settings. The section now describes them, notes that merge-on-read writes aren't intercepted yet (Support Native Iceberg MOR write #6240), and points at the production-quality epic ([EPIC] Production-quality native Iceberg writes #5649) and the default flip (Enable the split-operator plan and native Iceberg writes by default #5644). The closed draft feat: Native Parquet Iceberg Data File Writes In Comet #4487 is dropped.
• Iceberg Table Format V3 Support: native deletion vector reads landed in feat(iceberg): Iceberg table format V3: apply deletion vector on reads #5853, which replaced the closed draft feat: [DO NOT MERGE] Iceberg table format V3: native deletion vector reads, fall back for other V3 features #4887, and the upstream iceberg-rust epic has closed. The section now lists what still falls back: row lineage metadata columns, column default values, and the variant/geometry/geography/unknown types. The HDFS note referred to a scheme match in iceberg_scan.rs that has since moved to iceberg_common.rs. It now describes the supported storage backends without naming a file.
• Native Coverage for Codegen-Dispatched Expressions: the section said aggregate functions go through the codegen-dispatch bridge. The dispatcher only handles scalar expressions (CometBatchKernelCodegen rejects AggregateFunction), so incompatible aggregates fall back to Spark. The link now points at the expression reference, whose Implementation column lists the dispatched expressions; the compatibility guide it pointed at doesn't have that list.
• Java/Scala UDF Support: the section said Python UDFs always fall back. mapInArrow and mapInPandas have had experimental support since feat: Add experimental support for accelerated PyArrow UDFs #4234.
• TPC-H and TPC-DS Performance: the TPC-DS epic [EPIC] Improve performance of TPC-DS queries #858 is closed, so the section now points at Accelerate TPC-DS without any fallbacks to Spark #2551.
• Upstream Work in DataFusion: nearly every datafusion-spark function now has a Comet serde (Support expressions already implemented in datafusion-spark crate #4150); the remaining exception, json_tuple, is tracked in [Feature] Support Spark expression: json_tuple #3160.
• Spillable Hash Join: adds links to Add support for spilling in hash joins #2545 and to the upstream DataFusion design for spilling HashJoinExec ([EPIC] Spilling Hash Join — run any join in a bounded memory budget datafusion#24768).
• Delta Lake Support: describes the delta-spark contrib scan in review (feat: add native Delta Lake scan contrib module (page/row-group pruning) #5365) and the proposal to converge it with the delta-kernel-rs path (Converge the two Delta read paths into one plugin (clean architecture + performance) #5411). The dormant plain-table draft feat: support native Comet scan of plain Delta Lake tables #4669 is dropped because feat: add native Delta Lake scan contrib module (page/row-group pruning) #5365 supersedes its approach; I can add it back if we'd rather keep it listed.

The window, lambda, native Parquet write, and memory management sections are still accurate and are unchanged.

How are these changes tested?

This is a documentation-only change. prettier --check passes on the file, and every reference-style link it uses is defined, with no unused definitions.

[sunchao]

Summary

• Prior state and problem: The contributor roadmap described several implemented features as pending and linked to outdated tracking work.
• Design approach: Refresh capability descriptions and issue references in docs/source/contributor-guide/roadmap.md.
• Correctness / compatibility analysis: The revised claims agree with the relevant implementation and linked issue/PR states. Checked aggregate dispatch, Iceberg read/write gates, and Python support against Comet shims and relevant Spark sources across 3.4, 3.5, 4.0, 4.1, and experimental 4.2. Python acceleration remains limited to Spark 4.x.
• Key design decisions: Preserve experimental/default-off qualifications and distinguish implemented features from proposed work. The change adds no runtime abstraction or complexity.
• Implementation sketch: One documentation file changes, with 51 additions and 37 deletions. The entire base-relative diff was reviewed.
• Behavioral changes worth calling out: Contributor guidance changes. Execution behavior, Spark compatibility, and runtime overhead are unchanged.
• Suggested improvements: None meeting the P1/P2 bar. No introduced P1/P2 issues found within this review.

Reviewed head 47dea8e350adf711c220f18f31825ce3a77e2e0c against base 76493618769ff85e7a43e6f6c79f8b58d9e0b1ad. The PR is not a draft. The snapshot and live discussion contain no reviews, issue comments, inline comments, or review threads.

Routed skills: review-comet-pr. No sibling skill applies under its documentation-only routing rule.

Exact-head CI: Seven successful checks and fifteen skipped jobs, with no failures or unfinished checks. Runtime builds, Spark/Iceberg suites, and site deployment were skipped.

Validation: git diff --check, reference-definition checks, and local-link checks passed. Linked development statuses were verified. Prettier could not run because prettier/npx are unavailable, and Sphinx is not installed. No documentation build or runtime tests were run. The working tree remains unchanged.