[TRTLLM-13767][test] revert CUTLASS DSL 4.6.1 upgrade

[taianz-nv]

Summary

This draft validation PR reverts #17274 on top of the latest main to determine whether the pre-upgrade CUTLASS DSL 4.5.0 stack can still pass L0, and whether the regressions reported only escaped because the affected cases run in post-merge stages.

It restores the pre-upgrade CUTLASS DSL/kernel compatibility changes and dependency set, including CUTLASS DSL 4.5.0, FlashAttention 4.0.0b11, and the compatible quack-kernels resolution.

This is a test-only draft and is not intended to merge as-is.

Local checks

• pre-commit hooks passed
• poetry check --lock passed
• resolved the two expected security-scanning dependency/lock conflicts against current main

Validation

• Run the standard pre-merge L0 pipeline with /bot run.
• Run the affected MegaMoE and WAN cases separately on B200/B300 hardware to distinguish a real 4.5.0 result from missing pre-merge coverage.

[taianz-nv]

/bot run

[tensorrt-cicd]

PR_Github #68249 [ run ] triggered by Bot. Commit: 01495b0 Link to invocation

[tensorrt-cicd]

PR_Github #68249 [ run ] completed with state FAILURE. Commit: 01495b0
/LLM/main/L0_MergeRequest_PR pipeline #55685 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