Validate Attention projection and cache dimensions (#32946)
### Description
<!-- Describe your changes. -->
Shares Q/K/V, weight-shape, projection-size, and cache-length checks
between Attention prepacking and execution across CPU, CUDA, and native
WebGPU.
Uses checked packing/GEMM sizes, preserves pruned/unequal-V layouts and
shared prepacking, and supports omitted CPU projection bias. Adds
initializer-time, dynamic/unpacked, dimension-boundary, cache, and
graph-independent validation coverage; updates the operator
documentation.
Refreshed on main `397657de778d05b17c0249ba37c0b4290e1ab3b5`:
- Retains main's environment-based Android size allowance instead of
reverting it to the stale branch constant.
- Keeps both main's malformed-prepacking regressions and the new
unequal-hidden-size sharing test.
- Central validation preserves main's `"qkv_hidden_sizes values should
be positive"` diagnostic. Shared malformed-case expectations were
corrected after the first fresh native run exposed the mismatch.
- Includes main's independent CUDA GQA build integration fix.
### Motivation and Context
<!-- - Why is this change required? What problem does it solve?
- If it fixes an open issue, please link to the issue here. -->
Attention must enforce the same projection contract whether weights are
packed at initialization or supplied during execution. Status-based
validation rejects unsupported dimensions before unsafe packing/index
arithmetic and keeps cache capacity checks consistent between
implementations.
### Testing
Fresh Windows Release CPU/DirectML build with MSVC and Ninja passed.
MSBuild's CL task stalled without compiling; Ninja built the same
configuration successfully. Tests ran from its fresh executable
directory.
```powershell
.\onnxruntime_provider_test.exe --gtest_filter=ContribOpAttentionTest.*:ContribOpAttentionValidationTest.*
```
**57 passed, 3 explicit unavailable WebGPU/cuDNN skips, zero failures; 1
existing disabled test.** Includes the main prepacking diagnostics and
all three graph-independent helper cases. The initial run failed two
malformed-projection tests solely because their expected diagnostics did
not match the deliberately preserved main diagnostic; the corrected
expectations were rebuilt and this entire selector passed.
```powershell
.\onnxruntime_provider_test.exe --gtest_filter=QuantizeAttention*:*QAttention*
```
**20 passed, no skips**, verifying QAttention's shared validation.
```powershell
.\onnxruntime_mlas_test.exe --gtest_filter=FGemmPackB.SizeBounds
```
**1 passed, no skips**, including exact accepted/rejected packing-size
boundaries.
Historical earlier-head evidence: 3 isolated no-exception/minimal
validation cases and 601 packed-SGEMM cases passed. These configurations
were not rebuilt for the main reconciliation above; they are not
fresh-head minimal coverage.
**Final-head GPU execution is now verified**, not inferred from green
builds:
- [Linux CUDA test
job](https://github.com/microsoft/onnxruntime/actions/runs/37694054857/job/113070913707):
`InvalidProjectionSizesCUDA`, `InvalidWeightsRankCUDA`, and
`ProjectionSizeLimitCUDA` explicitly ran/passed; CTest **16/16 passed**.
- [Native Metal WebGPU
job](https://github.com/microsoft/onnxruntime/actions/runs/37694055050/job/113041537081):
`InvalidProjectionSizesWebGPU` and `InvalidWeightsRankWebGPU` explicitly
ran/passed; CTest **17/17 passed**.
Remaining CI: three Linux minimal/no-optional build jobs and Suggest
fixes are still genuinely queued without runners for more than eight
hours. They have not run and are not passes; provisioning attention is
requested without duplicating queued work. All other completed head
checks passed. Independent approval remains required.
---------
Co-authored-by: Gopalakrishnan Nallasamy <gopalakrishnan.nallasamy@microsoft.com>
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
