Refactor: extract compute_task_fanin / register_task_outputs from submit_task

[ChaoWao]

Summary

Pull the per-tensor dependency-computation logic of submit_task out
into a small header-only template in pto_dep_compute.h, shared between
the a5 and a2a3 tensormap_and_ringbuffer runtimes (they had identical
inline copies). Bit-equivalent refactor — no behavior change.

This is groundwork for an upcoming dep_gen feature (offline replay of
the orchestrator's dependency-discovery logic to capture a complete
dep graph without the race window in the in-record fanout[] snapshot).
Both submit_task and the future replay path will share the same
compute_task_fanin body, so dep computation has a single source of
truth.

What moved

New entry point	Original location
compute_task_fanin	STEP 3 in submit_task (per-tensor creator retention + tensormap lookup for INPUT/INOUT)
register_task_outputs	STEP 4 in submit_task (tensormap.insert for INOUT/OUTPUT_EXISTING)

DepInputs view struct decouples Arg's user-facing API from the
dep-computation interface so the future replay code can hand-roll a
DepInputs without faking an Arg.

What did not move (intentional)

STEP 1 (the explicit_dep loop) stays inline at the runtime call site.
Its last_task_alive short-circuit + unchecked slot lookup is subtly
different from STEP 3's slot-reuse check; unifying them would either
need two emit semantics or change transient behavior. The future replay
path will handle STEP 1 in a one-line emit loop of its own — duplication
is small and structural.

Performance

Emit is a template parameter, not std::function. Both runtime
(lambda capturing fanin_builder + sm_header) and future replay
(lambda capturing edge vector) instantiate at the call site and inline
all the way through. Do not replace with std::function — that would
add ~5 ns per call in the orch hot path. The header has a comment
warning future maintainers about this.

Cosmetic adds

• TensorTagMixin::tag_data() in src/common/task_interface/task_args.h
• Arg::explicit_deps_data() in
  src/a{5,2a3}/runtime/tensormap_and_ringbuffer/runtime/pto_types.h

Both are simple const T* accessors over already-public POD storage,
needed to build a DepInputs view without per-element copies.

Testing

• pre-commit run --files <changed> (clang-tidy, clang-format,
  cpplint, headers all pass on a5 and a2a3)
• a5sim: mixed_example (20 tasks, mixed AIC+AIV, RTOL=1e-3 golden
  check) PASSED
• a5sim: spmd_starvation (252 tasks scheduled stress) PASSED
• a5sim: spmd_basic (basic dispatch) PASSED
• a2a3sim: test_l3_dependency (5 tasks dep test) PASSED
• Linux CI green

[gemini-code-assist]

Code Review

This pull request introduces a new header file, pto_dep_compute.h, which provides template-based primitives for dependency computation. It refactors the submit_task method in PTO2OrchestratorState to utilize these new functions, improving code modularity. Additionally, helper methods explicit_deps_data and tag_data were added to the Arg and TensorTagMixin structures to support this refactoring. I have no feedback to provide as there were no review comments.