Fix: drop misaligned version guards in the L2 swimlane pipeline

[indigo1973]

deps.json and l2_perf_records.json both carry a "version" field, and
both have grown out of sync with the consumers that gate on it:

• deps.json: PR Refactor: a2a3 + a5 Tensor to strided (stride + start_offset) model #808 bumped v2 → v3 (strided Tensor: buffer_numel
  replaces raw_shapes, start_offset + strides[] replace multi-dim
  offset[]). swimlane_converter.load_deps_json kept the
  if version != 2 guard, silently rejected every fresh capture, and
  fell back to L2PerfRecord::fanout[] — losing the race-window edges
  that dep_gen replay exists to recover.

• l2_perf_records.json: its "version" field is actually L2PerfLevel
  (0=DISABLED, 1=AICORE_TIMING, ..., 3=SCHED_PHASES, 4=ORCH_PHASES),
  not a JSON schema version. swimlane_converter._print_verbose_data_info
  and sched_overhead_analysis.parse_scheduler_from_json_phases both
  gated phase output on version == 2 / version < 2, which is wrong
  on both ends: phase blocks only exist at level >= 3, so v3/v4 captures
  silently skipped phase summaries while v2 captures probed for fields
  that aren't there.

Bug fixes (semantics)

• swimlane_converter.load_deps_json: drop the deps.json version guard
  entirely. This consumer only reads edges[].pred / .succ — both stable
  across every deps.json schema to date. Malformed individual edges are
  still skipped per-iteration.
• swimlane_converter._print_verbose_data_info: drop the
  version != 2 short-circuit; .get() already handles missing
  phase fields gracefully.
• sched_overhead_analysis.parse_scheduler_from_json_phases: drop the
  version < 2 short-circuit; the if not phases_by_thread: return {}
  guard immediately below was already the right check.

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

• swimlane_converter: 3 inline "(version 2)" annotations → explicit
  "(l2_perf_level >= 3)" / "(>= 4)".
• sched_overhead_analysis: module docstring + run_analysis docstring
  • Pop fallback message replace "v2" / "version >= 2" wording with
    explicit l2_perf level.
• {a2a3,a5}/runtime/tensormap_and_ringbuffer/docs/profiling_levels.md:
  the "v2 JSON" paragraph now spells out which file, which field, and
  which level is required.
• dep_gen_replay.{h,cpp}: schema block + file brief + INFO log call
  out v3 fields (tasks[].args[], buffer_numel, start_offset, strides).
• deps_to_graph.py / docs/dfx/dep_gen.md / simpler_setup/tools/README.md
  / test_dep_gen.py: residual "v2" labels updated to v3. (The
  load-bearing version != 3 guard in deps_to_graph stays — that
  consumer reads strided-only fields.)

[gemini-code-assist]

Code Review

This pull request updates the dependency graph generation schema from version 2 (v2) to version 3 (v3), introducing a strided tensor representation. Key changes include adding an args array with detailed tensor slice geometry to tasks, replacing raw_shapes with buffer_numel in the tensors schema, and replacing simple offsets with explicit start offsets and strides for both consumers and producers in the edges schema. Downstream tools, documentation, and tests have been updated to support and validate the new v3 schema. There are no review comments, so I have no feedback to provide.