Add: dep_gen capture (SubmitTrace) for a2a3 tensormap_and_ringbuffer

[ChaoWao]

Summary

Capture every Orchestrator::submit_task call into a streaming
SubmitTrace ring on the a2a3 tensormap_and_ringbuffer runtime, drained
by a host collector to submit_trace.bin. Phase 1 of replacing #500 — the
host-side offline replay that reconstructs deps.json from these records
ships in a follow-up PR; this PR is intentionally scoped to the capture
path so it can land independently and be reviewed in isolation.

Motivation: today's swimlane L2PerfRecord::fanout[] is filled at the
producer's completion-record commit moment, so a fast producer that
finishes before a later consumer is submitted has its record sealed
without the new edge — those dep edges are silently lost. Capturing the
orch's submit-time inputs and replaying offline reconstructs the complete
logical graph because the replay can run with no eviction.

What lands in this PR

• Device structs — src/a2a3/platform/include/common/dep_gen.h:
  DepGenRecord (2240 B per submit: task_id, flags, arg_types[16],
  explicit_deps[16], tensors[16][128] as opaque blobs). Plus the
  PMU/L2Perf-style SPSC buffer family (FreeQueue / BufferState /
  ReadyQueueEntry / DataHeader). Single-instance (orch is one AICPU
  thread).
• AICPU writer —
  src/a2a3/platform/{include/aicpu,src/aicpu}/dep_gen_collector_aicpu.{h,cpp}:
  all-primitive interface (raw uint64 task_id, void* per-Tensor blob,
  uint64* explicit_deps) so the platform header stays runtime-agnostic.
  submit_task entry calls dep_gen_aicpu_record_submit gated on
  is_dep_gen_enabled(). Orch callsite static_asserts
  sizeof(Tensor) == DEP_GEN_TENSOR_SIZE and
  PTO2_MAX_EXPLICIT_DEPS == DEP_GEN_MAX_EXPLICIT_DEPS. Weak fallbacks
  let host builds link without the AICPU strong symbols.
• Host collector —
  src/a2a3/platform/{include/host,src/host}/dep_gen_collector.{h,cpp}:
  DepGenModule trait + DepGenCollector : ProfilerBase<...>, so the
  mgmt thread / buffer-pool manager / poll loop come from the unified
  profiling framework (Refactor(a2a3): decouple profiling from runtime, own it in platform #714). on_buffer_collected appends DepGenRecord
  values to submit_trace.bin; reconcile_counters cross-checks
  collected + dropped == device_total and returns clean/dirty so the
  future replay step can skip deps.json on incomplete traces.
• Device runner wiring (sim + onboard): DepGenCollector field,
  set_dep_gen_enabled() setter, init_dep_gen() helper,
  kernel_args.dep_gen_data_base plumbed through to AICPU, RAII cleanup,
  end-of-run flush + reconcile, make_dep_gen_path() helper.
• AICPU lifecycle: set_platform_dep_gen_base /
  set_dep_gen_enabled / dep_gen_aicpu_set_orch_thread_idx /
  dep_gen_aicpu_init / dep_gen_aicpu_flush /
  dep_gen_aicpu_finalize hooked alongside the existing PMU calls.
• CallConfig + Python: enable_dep_gen int32 flag, nanobind Python
  property (bool), conftest.py and scene_test.py --enable-dep-gen
  CLI option threaded through run_class_cases / _run_and_validate /
  _build_config. --rounds > 1 disables capture (matches the
  --enable-l2-swimlane pattern).
• a2a3sim test —
  tests/st/a2a3/tensormap_and_ringbuffer/dep_gen_capture/test_dep_gen_capture.py:
  re-uses the vector_example orchestration (5 submit_task calls), runs
  with --enable-dep-gen, then asserts submit_trace.bin size equals
  5 * sizeof(DepGenRecord) and spot-checks the first record's
  tensor_count == 3.

What does NOT land in this PR

• Host-side offline replay → deps.json
• swimlane_converter.py integration that prefers deps.json over the
  in-record fanout[]
• a5 port (the a5 profiling framework is the pre-Refactor(a2a3): decouple profiling from runtime, own it in platform #714 layout and needs
  separate wiring)

These follow in subsequent PRs once the capture path is reviewed.

Testing

• pre-commit run (clang-format, clang-tidy, cpplint, ruff,
  pyright) all pass
• a2a3sim spmd_sync_start baseline (no --enable-dep-gen)
  PASSED — wiring does not perturb the default path
• a2a3sim dep_gen_capture (with --enable-dep-gen) PASSED,
  post-run verification PASSED (submit_trace.bin has exactly
  5 records totalling 11200 bytes)
• Linux CI green

Notes

• profiler_base.h had to gain a one-line
  // NOLINTNEXTLINE(bugprone-crtp-constructor-accessibility)
  • comment: the new dep_gen_collector.cpp is the first translation
    unit in this checkout to pull the header into pre-commit's clang-tidy
    scope. Existing PmuCollector / L2PerfCollector use the same public
    ctor pattern; making it protected would force every derived class
    to declare friend ProfilerBase solely for ctor visibility.

[gemini-code-assist]

Code Review

This pull request implements the dep_gen (SubmitTrace) feature to capture orchestrator task submissions for offline replay and dependency graph reconstruction. The changes include new shared-memory protocols, AICPU-side recording, host-side collection, and integration into the Python and C++ runtime layers. Feedback was provided to improve the cache alignment of the tensors array in the DepGenRecord structure to meet performance and architectural standards.