[FlyDSL] Support runtime softmax scale in gfx950 FP8 FMHA

[gbyu-amd]

The gfx950 FlyDSL FP8 FMHA kernel currently fixes the attention scale to 1 / sqrt(D). Consumers using a different layer scale, such as YaRN, must fold the correction into a quantization descale, and the AITER dispatch rejects nondefault scales.

Add softmax_scale=None to the direct FP8 API, using the BF16 API's parameter name and default. Pass it as a runtime FP32 scalar and apply softmax_scale * q_descale * k_descale to the logits inside FMHA. Dense and varlen dispatch and the large-tensor batch-slicing path forward the scale. This needs no descale mutation or additional GPU kernel launch, and scale values do not enter the build-cache key. Output and LSE use the same scale.

The existing online softmax requires a positive, finite scale: it takes maxima and applies masking before scaling. The direct API rejects zero, negative and non-finite scales; dispatch returns None for those configurations. Supporting them would require a separate change to the softmax algorithm.

This PR targets main and contains only FMHA scale support, dispatch changes and tests. The review follow-up removes the redundant head_dim_runtime launch argument and shares softmax-scale validation between the dispatch gate and direct FP8 API.

Validation:

• Review follow-up (a859b8271): 24 CPU tests passed, 252 GPU tests skipped, with HIP/ROCR/CUDA device visibility disabled. Fake-tensor tests exercise the direct API and dense/varlen dispatch for default/custom/invalid scales; launch-boundary tests check default and custom scales in both run and compile argument preparation for D=128/192. No GPU execution or GPU compilation was performed. Black, Ruff and git diff --check pass.
• Prior validation after the rebase onto main: Black, Ruff and git diff --check pass; all 252 FMHA tests collect successfully with GPUs hidden (AITER_AOT_IMPORT=1). Collection is not a correctness run.
• At that rebase, git range-diff confirmed the feature patch was unchanged; all five changed files were byte-identical to the previously GPU-validated PR head. GPU tests were not repeated during the active TP8 workload.
• Added coverage checks nondefault scales, default/explicit-scale equivalence, repeated scale changes on a cached launcher, 12-head K3 shapes, dense/varlen, GQA, split-K, fully masked rows, zero-length KV entries, graph replay, dispatch and batch slicing.
• GPU validation before the main rebase, on a free gfx950 GPU with a fresh FlyDSL cache: 80 passed, 1 skipped (the cross-device test requires a second visible GPU). This includes the custom-scale correctness, LSE, dispatch and graph-replay tests above.
• ATOM integration after replacing descale folding with direct softmax_scale forwarding: 112 tests passed.
• Cached-chunk GPU traces for Q=70, H=12, D=192, Dv=128 and KV=16,384 / 11,008 confirm one FMHA kernel per call, with no added scale-preparation launch. Warmed CUDA-graph measurements were 278.99 / 190.98 microseconds per call (7 rounds of 100 replays); these are absolute measurements, not a speedup claim for this change.

[binding7012]

LGTM