[Feature] l0_swimlane: restore real tensor payloads so metadata-driven kernels are replayable

[ChaoWao]

Summary

l0_swimlane reconstructs a task's args from a level-3 dump — descriptor
metadata only, no payload (l0_swimlane.py:264 runs --dump-args 3). Tensor
buffers are therefore aclrtMemset to 0 in the replay, and --set-arg is the
only way to give one content: it writes a single integer into every element
(l0_swimlane.py:643-652).

That covers a scalar or a uniform control tensor, but not a kernel whose control
input is a struct. Proposal: let the tool restore real bytes for selected
tensors from a --dump-args 2 dump, which already captures them.

Motivation / Use Case

examples/a2a3/tensormap_and_ringbuffer/qwen3_14b_decode is currently
unreplayable for this reason, and the l0 doc's "real SPMD workload" category has
no representative because of it (see #1529).

Its attention is a CANN FusedInferAttentionScore extern that reads a
FAInferTilingData struct out of a UINT8 metadata buffer filled at runtime by
paged_attention_tiling_cce. Uniform fill cannot synthesize a coherent struct —
and UINT8 is in INTEGER_DTYPES, so --set-arg does not even refuse it
(l0_swimlane.py:104, :1231); it just produces nonsense. With metadata
zeroed, fai_body.hpp reads tiling->needCoreNum == 0 and takes the degenerate
branch, so the trace comes out unrepresentative rather than empty — the
worse of the two failures, because it looks like a result.

Being an extern is not the blocker: the replay feeds kernel_entry and does not
care whether the source was generated or hand-written. The blocker is purely
that the tool cannot deliver a structured buffer. Any kernel taking its work
distribution from a runtime-built descriptor hits the same wall, which is the
normal shape for CANN operator wrappers.

Proposed API / Behavior

Add an opt-in that sources a tensor's initializer from a payload-carrying dump
rather than a constant, e.g.:

$L0a --func-id 11,12 --restore-arg 6 --test $E/qwen3_14b_decode/test_qwen3_14b_decode.py

• --restore-arg SLOT (repeatable): emit that tensor's captured bytes into the
  generated workspace and aclrtMemcpy them in, instead of aclrtMemset(0).
• Implies a level-2 dump for the run (or errors clearly if reusing a level-3
  dump that has no .bin).
• Composes with --set-arg; conflicting slots should be rejected.

Scope note: only worth restoring for control inputs. Restoring every tensor
would drag the full working set (weights, KV pool) into the workspace, which is
why level 3 exists in the first place (#1020).

Alternatives Considered

• --set-arg on the metadata buffer — what exists today. Cannot express a
  struct; silently yields a degenerate trace.
• Hand-author the replay workspace (the insight-trace path) — works, but
  gives up exactly what l0_swimlane was built for ([Feature] Generate MindStudio Insight replay traces from kernel args dump #836): real captured args
  with no shape-guessing, and it must be redone whenever the tiling layout moves.
• Synthesize the tiling in the replay by calling the tiling kernel first —
  turns a single-op replay into a two-op one and reintroduces the ordering the
  replay recipe deliberately flattens.
• Leave the category unrepresented — the status quo after Docs: correct two claims left behind by the qwen3 re-harvest #1529, and the
  reason for filing this.

Additional Context

• Gap documented in docs/dfx/l0-swimlane-profiling.md (§ "real SPMD workload"),
  updated by Docs: correct two claims left behind by the qwen3 re-harvest #1529.
• Dump levels: docs/dfx/args-dump.md — level 2 is "full — every task's tensor
  inputs/outputs", level 3 is "full, JSON only … no payload captured".
• Related (all closed, this builds on them): [Feature] Generate MindStudio Insight replay traces from kernel args dump #836 (Insight replay from an args
  dump), [Feature] Add per-kernel dispatch args dump for Insight Trace #837 (per-kernel dispatch args dump), [Feature] Add JSON-only tensor-dump level (metadata, no .bin payload) #1020 (the JSON-only level this
  proposal complements).
• Observed at d97b932f.