[TRTLLM-12339][feat] enable TRTLLM cross attention backend

[cascade812]

Description

Split out the attention operator and TRTLLM attention backend changes from #13919 to reduce frequent conflicts with main and make CI validation easier for this smaller, self-contained scope.

This PR intentionally keeps the change self-contained:

• wires thop.attention and its nanobind signature for cross-attention and relative-attention-bias inputs
• enables the TRTLLM backend path for cross-attention metadata, including Q padding, cross K/V forwarding, and beam-width handling
• makes trtllm-gen decline cross-attention so cross requests use the THOP path
• adds only the small backend forward-args fields required by the TRTLLM backend

No module, executor, model, or LLM API caller changes are included in this split.

Summary by CodeRabbit

• New Features
  • Added cross-attention support with optional cross-key-value tensor inputs.
  • Added optional relative attention bias with configurable maximum distance parameter.

[cascade812]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #54082 [ run ] triggered by Bot. Commit: 46ef1af Link to invocation

[tensorrt-cicd]

PR_Github #54082 [ run ] completed with state SUCCESS. Commit: 46ef1af
/LLM/main/L0_MergeRequest_PR pipeline #43166 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[cascade812]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #54328 [ run ] triggered by Bot. Commit: 3dfaeac Link to invocation

[tensorrt-cicd]

PR_Github #54328 [ run ] completed with state FAILURE. Commit: 3dfaeac
/LLM/main/L0_MergeRequest_PR pipeline #43399 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[cascade812]

/bot run

[tensorrt-cicd]

PR_Github #54380 [ run ] triggered by Bot. Commit: 3dfaeac Link to invocation

[tensorrt-cicd]

PR_Github #54380 [ run ] completed with state SUCCESS. Commit: 3dfaeac
/LLM/main/L0_MergeRequest_PR pipeline #43450 completed with status: 'SUCCESS'

CI Report

Link to invocation

[brb-nv]

LGTM.