Preserve distinct epsilons in GQA pre-norm fusion (#33363)
### Description
Preserve separate Q and K normalization epsilons when fusing into
GroupQueryAttention. Add optional `q_norm_epsilon` and `k_norm_epsilon`
attributes, each falling back to the existing `qk_norm_epsilon`, and
carry them through WebGPU and CUDA normalization paths. Equal-epsilon
fusion retains the shared attribute; mismatched epsilons now fuse with
both original values preserved. Update the generated attribute
documentation and regression tests.
### Motivation and Context
The fusion previously accepted Q/K epsilon differences up to 1e-9 and
then applied Q's epsilon to K. Preserving both values fixes the
numerical error while allowing fusion for all otherwise-supported
positive finite epsilon pairs.
Validation on Windows with fresh builds from this branch:
| Build | Fusion tests | QK-normalization operator tests |
| --- | --- | --- |
| WebGPU | 21 passed, 4 CUDA-only skips | 5 passed, 6 CUDA-only skips |
| CUDA 13.1 / cuDNN 9.19 | 21 passed, 4 WebGPU-only skips | 7 passed, 4
WebGPU-only skips |
All 36 distinct focused tests passed across the two builds. New GPU
fusion comparisons disable CPU fallback and check attention and cache
outputs for decode and prefill. Tests cover independent overrides, the
legacy fallback, invalid values, and stale attributes. The issue's
attached model fuses with both epsilons preserved and matches all three
unfused WebGPU outputs exactly (maximum absolute differences: 0, 0, 0).
CUDA validation targeted SM120 with FlashAttention and memory-efficient
attention disabled; those paths were not numerically validated.
Changed-line clang-format 20.1.8, `git diff --check`, and the GQA
attribute documentation check passed. No performance claim is made.
Fixes #32930.
