[Feature] Track reader accesses in TensorMap and derive WAR dependencies automatically

[ChaoWao]

Summary

Extend TensorMap from a producer-only map into an outstanding-access tracker so automatic dependency generation derives write-after-read (WAR) edges for overlapping tensor regions. Pure INPUT readers must remain concurrent with each other, while a later INOUT or OUTPUT_EXISTING writer automatically waits for every overlapping live reader.

Motivation / Use Case

Issue #1306 demonstrated a deterministic correctness failure for a loop-carried buffer:

W0: add_inout(buf)
R0: add_input(buf)
W1: add_inout(buf)

The current TensorMap records only writers, so it derives W0 -> R0 (RAW) and W0 -> W1 (WAW), but misses R0 -> W1 (WAR). W1 may overwrite buf while R0 is reading it, causing silent wrong results.

Promoting readers to INOUT serializes independent readers. Requiring explicit dependencies makes every frontend/user responsible for detecting aliasing WAR patterns. Automatic scopes should preserve correctness without losing reader concurrency.

Related: #1306 and hw-native-sys/pypto#2058.

Proposed API / Behavior

Keep the public orchestration API unchanged and enhance TensorMap internally.

1. Add an access kind (READER/WRITER) to PTO2TensorMapEntry; generalize producer_task_id into the task owning the access.
2. After fanin calculation, register automatic-scope, non-manual_dep INPUT arguments as reader entries. Continue registering INOUT and OUTPUT_EXISTING as writers.
3. Apply access-specific lookup rules:

INPUT:
    depend on overlapping writers (RAW)
    never depend on readers
    register reader

INOUT:
    depend on overlapping writers (RAW/WAW)
    depend on overlapping readers (WAR)
    retire fully covered old accesses
    register writer

OUTPUT_EXISTING:
    preserve existing write-only semantics for old writers
    depend on overlapping readers (WAR)
    retire fully covered old accesses
    register writer

4. Reuse existing overlap and per-task cleanup semantics. A fully covering writer may remove an old access after wiring its dependency; partially overlapping entries must remain. Reader entries become stale when their task retires.
5. Host reads should wait only for writers. Host writes, including set_tensor_data(), should also wait for overlapping readers.
6. Replace output-only entry counting with access-entry counting. Re-evaluate entry-pool/bucket sizing, arena memory, pool back-pressure, bucket-chain length, and fanin spill capacity.
7. Keep runtime and replay behavior equivalent across a2a3 tensormap_and_ringbuffer, a2a3 host_build_graph, a5 tensormap_and_ringbuffer, and dep_gen_replay/DFX paths.

A correctness-first implementation may use one mixed chain with access-kind filtering. If reader-heavy profiling shows a significant RAW lookup regression, split reader/writer bucket indexes while sharing the entry pool.

Acceptance criteria

• W0(inout X) -> R0(input X) -> W1(inout X) contains R0 -> W1.
• Multiple pure readers remain mutually independent.
• A writer depends on every overlapping live reader, with fanin deduplication.
• Disjoint regions create no WAR edge.
• Partial overlap preserves the reader entry; full coverage may retire it after wiring.
• Retired readers create no stale dependency.
• Cross-ring WAR works.
• set_tensor_data() waits for a pure INPUT reader of the target region.
• The [Bug] Task dep-gen misses WAR edge: reader task not ordered before a later aliasing inout writer #1306 loop-carried reproducer passes on supported variants.
• Chain length, pool high-water mark, orchestration latency, and fanin spill usage are measured before default enablement.

Alternatives Considered

• Promote readers to INOUT: creates the edge but serializes readers as WAW.
• Require add_dep: precise, but omitted detection silently corrupts results.
• Detect only in PyPTO: useful for compatibility, but does not protect other orchestration producers.
• Start with separate maps: cleaner reader-heavy performance but a larger initial layout change than a mixed access-kind chain.

Additional Context

Relevant implementation areas:

• src/{arch}/runtime/{variant}/runtime/pto_dep_compute.h
• src/{arch}/runtime/{variant}/runtime/pto_tensormap.h
• src/{arch}/runtime/{variant}/runtime/pto_orchestrator.cpp
• src/{arch}/runtime/{variant}/runtime/pto_runtime2.cpp
• src/{arch}/runtime/{variant}/host/dep_gen_replay.cpp

Current commit reviewed: cddf2b6cb38e2f9bcc4739b5f888feaf33129675.