Add complete GQA workspace route recipes (#32454)
## Summary
Stacked on #32446 and #32453, complete the graph-free CUDA
GroupQueryAttention selected-route workspace model:
- add CUTLASS memory-efficient Attention K/V expansion and optional FP32
accumulator recipes;
- add the unfused aligned Q/Y/FP32-QK/softmax workspace recipe;
- compose preparation and exactly one selected backend into one checked
256-byte-aligned operator root;
- preserve effective staged KV-cache capacity for MEA;
- report cuDNN workspace as unavailable rather than treating its
allocator-based graph workspace as zero.
## Scope
This PR creates concrete selected-route roots. It does not enumerate
routes reachable under dynamic bounds, add Level-1/Level-2 framework
adapters, query cuDNN graphs, change runtime allocation topology, or opt
into #32071 planning.
## Validation
- CUDA provider test target built successfully.
- 213 CUDA internal tests passed; 2 unrelated Lean Attention tests
skipped.
- 56 GQA workspace tests passed, including MEA, unfused, complete-route
composition, XQA/Flash parity, and cuDNN-unavailable behavior.
- 73 related GroupQueryAttention runtime tests passed; 12 WebGPU-only
tests skipped because WebGPU was unavailable.
Tracking: #29775
Dependencies: #32446 and #32453
> This draft temporarily targets `main` because GitHub cannot use a
fork-only branch as the base of an upstream PR. Its diff will reduce to
this commit after the preceding stacked PRs merge.
---------
Co-authored-by: copilot-swe-agent[bot] <198982749+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Copilot-Session: b24d04bb-12ab-41a0-8941-76eb9769005d
