feat(tilelang): Add End2End tilelang profile backend and example

[zihaomu]

Summary

This PR is a follow-up to #215 and completes the TileLang profiling path in GEAK.

#215 adds the basic TileLang backend support, while this PR focuses on the missing TileLang-specific profiling workflow. With this change, GEAK can run TileLang workloads through an end-to-end correctness/profile/benchmark flow, collect profiling signals in ROCm environments, and use those signals to guide optimization.

This PR also adds a controlled TileLang GEMM optimization example. In our experiment, the TileLang GEMM example achieved about 1.5x speedup with GEAK.

What changed

• Add end-to-end TileLang profiling support for GEAK.
• Add TileLang ROCm profiling policy and fallback handling.
• Add legacy rocprof profiling support for TileLang workloads.
• Harden TileLang profile degradation behavior so profiling failures do not break the whole optimization flow.
• Fix structured command profile targets for TileLang profiling.
• Add a controlled examples/tilelang_gemm task, including:
  • target TileLang GEMM kernel
  • official reference kernel
  • correctness/profile/benchmark harness
  • GEAK-compatible task/config files
  • documentation for running the example manually and with GEAK
• Keep profile guidance tests independent of the TileLang backend availability.

Dependency

This PR should be reviewed and merged after #215.

#215 introduces the TileLang backend integration. This PR builds on top of that work by adding the complete TileLang profiling support and an end-to-end TileLang GEMM optimization example.

Validation

The new TileLang GEMM example supports:

  cd examples/tilelang_gemm
  python test_tilelang_gemm_harness.py --correctness
  python test_tilelang_gemm_harness.py --full-benchmark --iterations 2 --warmup 1

  Expect output：

  CORRECTNESS: PASS
  GEAK_RESULT_LATENCY_MS=...