Refactor: migrate legacy tests to @scene_test format

[doraemonmj]

Summary

Migrate all legacy golden.py + kernel_config.py tests under a2a3/tensormap_and_ringbuffer/ to the unified @scene_test / SceneTestCase format (single test_*.py per example).

• 10 synthetic examples moved from examples/ → tests/st/ with new test_*.py
• 3 tests/st/ tests converted in-place (benchmark_bgemm, paged_attention_unroll, explicit_fatal kernels)
• 2 dual-directory merges (batch_paged_attention, paged_attention): small-scale sim cases + production-scale hardware cases → single test_*.py
• batch_paged_attention kernels unified: C++ typename T template parameter + 3-way runtime dispatch (fp16/bf16)
• Include path fix: kernel_compiler.py adds runtime_common_dir to resolve intrinsic.h
• Zero legacy files remaining: all golden.py / kernel_config.py deleted

Migration Table

Location Changes

#	Case	Before	After	Type
1	spmd_basic	examples/	tests/st/	synthetic → st
2	spmd_multiblock_aiv	examples/	tests/st/	synthetic → st
3	spmd_multiblock_mix	examples/	tests/st/	synthetic → st
4	spmd_starvation	examples/	tests/st/	synthetic → st
5	spmd_sync_start	examples/	tests/st/	synthetic → st
6	spmd_sync_start_aiv	examples/	tests/st/	synthetic → st
7	spmd_sync_start_edge	examples/	tests/st/	synthetic → st
8	spmd_sync_start_stress	examples/	tests/st/	synthetic → st
9	mixed_example	examples/	tests/st/	synthetic → st
10	multi_round_paged_attention	examples/	tests/st/	synthetic → st
11	benchmark_bgemm	tests/st/	tests/st/	in-place convert
12	paged_attention_unroll	tests/st/	tests/st/	in-place convert
13	batch_paged_attention	examples/ + tests/st/	examples/	dual-dir merge
14	paged_attention	examples/ + tests/st/	examples/	dual-dir merge
15	bgemm	examples/	examples/	in-place convert
16	paged_attention_ringbuffer	examples/	examples/	in-place adjust

Parameters (all unchanged)

#	Case	RTOL/ATOL	thread_num	block_dim	Cases
1	spmd_basic	0/0	4	24	Case1: {}
2	spmd_multiblock_aiv	0/0	4	24	Case1: {}
3	spmd_multiblock_mix	0/0	4	24	Case1: {}
4	spmd_starvation	0/0	4	24	Case1: {}
5	spmd_sync_start	0/0	4	24	Case1: {}
6	spmd_sync_start_aiv	0/0	4	24	Case1: {}
7	spmd_sync_start_edge	0/0	4	24	Case1: {}
8	spmd_sync_start_stress	0/0	4	24	Case1: {}
9	mixed_example	1e-3/1e-3	4	3	case1: {num_iters:4}, case2: {num_iters:1}
10	multi_round_paged_attention	1e-2/1e-2	4	24	Case1: {batch:1, ...} manual=True
11	benchmark_bgemm	1e-3/1e-3	4	24	Case0-4 (5 cases)
12	paged_attention_unroll	1e-3/1e-3	4	24	Case1-3: prod bf16, manual=True
13	batch_paged_attention (small)	1e-2/1e-2	4	24	Case1-3+VarSeq2+VarSeq4: fp16
13	batch_paged_attention (large)	1e-3/1e-3	4	24	Case1-3: bf16, manual=True, RING_HEAP=1GB
14	paged_attention	1e-2/1e-2	4	24	Case1-2+VarSeq2+VarSeq4: bf16
15	bgemm	1e-3/1e-3	4	3	default: {4x4x4 grid}
16	paged_attention_ringbuffer	1e-3/1e-3	4	24	ringbuffer_stress: {batch:32, bf16}, env vars

Key Changes

Change	Description
Format conversion	golden.py + kernel_config.py → single test_*.py (SceneTestCase)
Explicit platforms	sim-compatible: ["a2a3sim", "a2a3"], hardware-only: ["a2a3"]
batch_paged_attention kernel merge	Two kernel sets → one with typename T + 3-way runtime dispatch
batch_paged_attention class merge	examples/ small(fp16) + tests/st/ large(bf16) → two classes in one file
paged_attention dual-dir merge	examples/ small + tests/st/ large → examples/ small only
paged_attention_ringbuffer	Kernel path → ../batch_paged_attention/kernels (reuse merged kernels)
PA generate_args	Removed .flatten(), keep original tensor shapes
PA compute_golden	Generic args.specs traversal instead of manual reshape
Include path fix	kernel_compiler.py adds src/{arch}/runtime/{rt}/common/ for intrinsic.h
Legacy cleanup	All golden.py / kernel_config.py deleted, zero remaining

[gemini-code-assist]

Code Review

This pull request refactors the Batch Paged Attention implementation and its test suite, introducing production-scale chunked batching and runtime dispatch for multiple tile configurations and data types. It consolidates test logic into SceneTestCase classes and adds extensive functional and stress tests for SPMD, BGEMM, and mixed AIC/AIV operations. Feedback points out an unnecessary .flatten() call in the multi-round paged attention test, which is inconsistent with the PR's objective of maintaining original tensor shapes and complicates the golden result computation.