[debug] Instrument TLE cluster GEMM CI failure

[sunnycase]

Temporary CI debug PR for the 04-cluster-gemm.py Hopper failure in PR #649.

Changes:

• Print runtime/cache/autotune-candidate diagnostics when FLAGTREE_CI_DEBUG_CLUSTER_GEMM=1.
• Run the suspected BK=256 remote config under CUDA_LAUNCH_BLOCKING=1 before the original full default run in the triton_v3.6.x Hopper workflow.

This PR is intended for diagnostics only and should be closed after collecting logs.

[sunnycase]

CI debug summary for PR #649:

• Original PR [TLE] Feature/extrac tile strides new #649 Hopper rerun still fails in python/tutorials/tle/04-cluster-gemm.py with CUDA error: misaligned address at _autotune_config -> torch.cuda.synchronize() immediately after baseline triton_gemm autotune.
• Debug run with CUDA_LAUNCH_BLOCKING=1 passed. Candidate 1 for remote autotune was KernelConfig(bm=32, bn=512, bk=256, num_warps=4, num_stages=2) and completed launch/sync/bench successfully.
• Debug run without CUDA_LAUNCH_BLOCKING, but with per-candidate print(..., flush=True), also passed; all 16 remote candidates completed.

Current inference: the failure is timing-sensitive / host-side perturbation-sensitive, not a deterministic bad config. The first remote candidate remains the likely failing launch in the uninstrumented script, but adding prints is enough to hide it on the Hopper runner.