fix(sampling): acquire PDL inputs before softmax reads - #1682

Problem:
The fused, map, and reduce softmax variants read per-request temperature
before griddepcontrol.wait. Under programmatic dependent launch, a
consumer could therefore observe stale producer data and emit all-NaN
probabilities. A delayed-producer control reproduced the defect with PDL
on; the same control was finite and exact with PDL off.

Correction:
Move the dependency wait ahead of every dependent temperature read and
add a compiler memory clobber. Preserve softmax arithmetic, dtype,
temperature semantics, and the existing PDL launch policy.

Validation on GB200 (recorded before this message-only amendment):
- Original native fails the delayed-producer regression; corrected code
  passes eager and CUDA-graph producer/consumer tests.
- Final suite: 120 passed, 36 strict expected failures for an unrelated,
  pre-existing zero-temperature padded-lane NaN case.
- Serialized old/new comparison: 60 of 60 outputs bitwise identical,
  including the pre-existing NaN cases.
- PDL stress control: 16 of 16 finite and exact across four GPUs.
- Reduced-output graph replay with the separate request-local RNG fix:
  48 of 48 passed. This is not a full model or serving evaluation.
- Native sm100a disassembly confirms acquire precedes temperature load.
- pre-commit run --all-files: passed again before this amendment.

Performance:
No clean latency or throughput comparison was collected; no speedup is
claimed. This is a synchronization correctness fix, not PDL disablement.
The original unobserved replay failure is not assigned a proven cause,
and the unrelated zero-temperature bug remains open.

Signed-off-by: yechank <161688079+yechank-nvidia@users.noreply.github.com>