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:

04150de feat(relu): add the cuTile ReLU backward kernel
a71eafe Stop TileCpp compiler descendants and support opt-in compile deadlines
7d26e3a Bound CUDA Graph calibration duration for long kernels
24f63d4 Honor Test OP autotune disable in cuTile matmul paths
d9ba8fb test(matmul): skip FP64 large matmul OOM on B20X (sm120)
4b0b10a perf(mla_decoding_split_kv cutile): shape-aware TILE_H
2aab72d Add a compilable smaller grouped-GEMM candidate on SM80
36a99cd Fix TileCpp attention sink CUDA Graph capture and dynamic offsets
7b6fd7f Fix cuTile-RS CUDA Graph launches and contain FFI panics
2c6cc16 Fix benchmark capture lifetime, profiling, and comparison memory
ac22157 style(tilecpp): reflow the nvcc invocation
f4cddb6 Unify cold-cache CUDA Graph measurement and autotuning

CI Configuration

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

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