perf(zccache) Phase 2: route per-compile dispatch through embedded ZccacheService

[zackees]

Sub-issue of #789. Depends on Phase 1.

Scope

With FBUILD_ZCCACHE_EMBEDDED=1, every per-TU compile goes through
ZccacheService::compile(...) directly instead of zccache wrap … child
processes. Wrapper path remains intact for default builds.

Deliverables

• New embedded dispatch path in compile_many.rs / compile_exec.rs (or
  the equivalent call site under fbuild-build) selected by the
  CompileBackend enum landed in Phase 1.
• Preserve cache-key compatibility: the wrapper-mode helpers
  (compile_cwd_from_output, normalize_flags_for_compile_cwd,
  path_arg_for_compile_cwd, UNC \?\ stripping) must be applied in
  the embedded path too — or moved inside the embedded service if it
  takes raw args. Goal: warm caches built by wrapper mode keep hitting
  after a flag flip.
• Parallel-compile contract: confirm ZccacheService is Sync and
  produces no single-mutex bottleneck under the existing
  compile_many.rs fan-out. If it does, add a handle-pool layer.
• ensure_running (the zccache start spawn at Daemon teardown hardening: process containment + socket lingering + console events #32) becomes a no-op in
  embedded mode — no child daemon needed.

Acceptance

• With FBUILD_ZCCACHE_EMBEDDED=1, building tests/platform/uno
  produces a hot cache on rebuild and no zccache.exe child process
  is spawned (verified via process snapshot inside the test).
• Wrapped vs embedded produce byte-identical object files for
  tests/platform/uno + tests/platform/esp32dev, or the diff is
  explained and documented.
• Cold + warm wall-clock timings published in the PR for both
  platforms — quantifies the wrapper-spawn overhead this removes.
• compile_many.rs parallelism unchanged: same -j scaling under
  embedded as under wrapped.
• All existing zccache.rs tests still pass under wrapped mode.