[WIP] Fix: clean up "version" fields in L2 swimlane / dep_gen JSON

[indigo1973]

deps.json and l2_perf_records.json both carried a "version" field that consumers were getting wrong:

• deps.json bumped v2 → v3 in Refactor: a2a3 + a5 Tensor to strided (stride + start_offset) model #808 but swimlane_converter still guarded on version != 2, silently rejected every fresh capture, and fell back to L2PerfRecord::fanout[] — losing the race-window edges dep_gen replay exists to recover.

• l2_perf_records.json's "version" was never a schema version — the producer writes L2PerfLevel (1..4) there. Misreading it caused swimlane_converter._print_verbose_data_info and sched_overhead_analysis to short-circuit on version != 2 / < 2, while phase blocks only exist at level >= 3.

deps.json — drop the "version" field
Producer no longer emits it; deps_to_graph drops its version guard (same release as the producer; KeyError is a clearer failure than a synthetic guard); test_dep_gen + test_dep_gen_chain drop the version assertion; dep_gen_replay.{cpp,h}, docs/dfx/dep_gen.md, and tools/README.md drop the v2/v3 schema labels. dep_gen.md §4 example JSON, fields table, and §5 arg-row description are rewritten against the current strided-Tensor producer (buffer_numel replaces raw_shapes; start_offset + strides[] replace multi-dim offset[]).

l2_perf_records.json — rename "version" → "l2_perf_level"
Producer (a2a3 + a5) writes the new name; swimlane_converter.
read_perf_data + verbose print + _swimlane_validate.py follow.
Both misaligned short-circuits removed: load_deps_json's guard
outright (only reads edges[].pred / .succ, stable across every
schema), _print_verbose_data_info's version != 2, and
parse_scheduler_from_json_phases's version < 2 (the
if not phases_by_thread check below was already correct).

Doc / comment fallout — keep code, comments, and docs in sync per .claude/rules/doc-consistency.md:

• swimlane_converter / sched_overhead_analysis / profiling_levels.md (a2a3 + a5): "(version 2)" / "v2 JSON" wording → explicit l2_perf_level >= N.
• 6 scheduler comments (a2a3 + a5: scheduler_dispatch.cpp, scheduler_cold_path.cpp, scheduler_types.h) describing "v2 JSON dispatch records" now point at dispatch-phase records in aicpu_scheduler_phases[].
• tools/README.md Input File Format example is now a real two-task slice; the stale stub had wrong field names and missed ring_id / dispatch_time_us / finish_time_us. Troubleshooting "Unsupported version" entry renamed to "Unsupported l2_perf_level".
• docs/dfx/l2-swimlane-profiling.md: level table, per-task table, phase-record table, and prose all corrected — start_time → us, core_type 0/1 → "aic"/"aiv", phase_id SCHED* → lowercase phase strings, plus ring_id / pop_hit / loop_iter vs submit_idx / core_to_thread[] additions.

[gemini-code-assist]

Code Review

This pull request updates the profiling and dependency graph schemas, replacing the legacy "version" fields in deps.json and l2_perf_records.json with a descriptive l2_perf_level and a strided-tensor representation. The changes update the documentation, downstream analysis tools, tests, and C++ collectors/replay logic to use buffer_numel, start_offset, and strides instead of raw shapes and offsets. One issue was identified in dep_gen_replay.cpp where strides are implemented and serialized as unsigned integers (uint32_t) despite being documented as signed int32 arrays, which could cause serialization issues for negative strides.

[indigo1973]

The modifications have been pushed to PR856.