Unify cold-cache CUDA Graph measurement and autotuning & style(tilecpp): reflow the nvcc invocation & other updates

[hannahli-nv]

Description

Update codes.

This PR contains 12 new commit(s).

Commits included:

e69e27d feat(relu): add the cuTile ReLU backward kernel
e1bf5e7 Stop TileCpp compiler descendants and support opt-in compile deadlines
4b81c3f Bound CUDA Graph calibration duration for long kernels
846f3fc Honor Test OP autotune disable in cuTile matmul paths
fbf039d test(matmul): skip FP64 large matmul OOM on B20X (sm120)
c38e87a perf(mla_decoding_split_kv cutile): shape-aware TILE_H
52af9e6 Add a compilable smaller grouped-GEMM candidate on SM80
943483d Fix TileCpp attention sink CUDA Graph capture and dynamic offsets
00f3ddd Fix cuTile-RS CUDA Graph launches and contain FFI panics
61c1cca Fix benchmark capture lifetime, profiling, and comparison memory
3b8c8c6 style(tilecpp): reflow the nvcc invocation
112b858 Unify cold-cache CUDA Graph measurement and autotuning

CI Configuration

config:
  build: true
  # valid options are "ops" and "benchmark"
  test: ["ops", "benchmark"]

Checklist

• Code formatted and imports sorted via repo specifications (./format.sh)
• Documentation updated (if needed)
• CI configuration reviewed

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[hannahli-nv]

/ok to test e69e27d