Enable bounded non-windowed GQA workspace estimation

[Chi Lo (chilo-ms)]

Description

Follow-up to #32617 for non-windowed CUDA GroupQueryAttention workspace estimation.

• add ep.cuda.gqa_workspace_max_total_sequence_length as an explicit positive bound for the runtime total_sequence_length scalar used by Level-2 declaration;
• size non-windowed present-cache/backend workspace from the larger of the scalar bound and past-cache capacity;
• model MayInplace conservatively by including the valid one-sided past/present alias case and its full past-tensor preservation buffer; and
• keep XQA and Flash fast decode out of the partial-alias case because those runtime routes require both cache pairs to alias.

The Level-1 node adapter remains unavailable for non-windowed inputs because it cannot access session configuration. This PR changes estimation and declaration only; it does not consume a planned workspace root or change runtime allocation topology.

Validation

• Release CUDA provider target built successfully on RTX 5090 / CUDA 13.3.
• Changed-file lintrunner and git diff --check passed.
• Added direct estimator, aggregate, and kernel-declaration coverage in group_query_attention_workspace_estimate_test.cc.
• The updated CUDA-internal test source compiled locally. Running the module was blocked by the existing Windows onnxruntime_providers_cuda_ut CMake dependency cycle/module-loading setup when internal tests are enabled.

Tracks #29775.

[Copilot]

Copilot review overview

🟡 Changes recommended

The new kernel-declaration test disables every backend route compatible with its model and will fail.

Get a fresh assessment by requesting another Copilot review.

Review effort: Balanced
Findings: 1 · 1

Open (2)

• Test expectation is unreachable with XQA disabled · New
• Clarify option is required for non-windowed declarations · New

What changed in this PR

Adds bounded Level-2 workspace estimation for non-windowed CUDA GroupQueryAttention.

Changes:

• Adds a session-configured total-sequence-length bound.
• Models one-sided cache alias preservation.
• Adds estimator, aggregation, declaration tests, and documentation.

File	Description
group_query_attention.h	Stores the configured bound.
group_query_attention.cc	Parses and applies the bound.
group_query_attention_workspace_estimate.h	Extends estimator configuration.
group_query_attention_workspace_estimate.cc	Builds non-windowed bounds.
group_query_attention_workspace_bounds.h	Adds past-capacity and alias metadata.
group_query_attention_workspace_bounds.cc	Aggregates partial-alias workspace routes.
group_query_attention_workspace_estimate_test.cc	Tests estimation and declaration behavior.
onnxruntime_session_options_config_keys.h	Defines the new session option.
attention_workspace_estimation.md	Documents non-windowed estimation.

---

💡 Configure MCP servers for context-aware, tailored reviews. Learn more in the docs.