Disable sam_fast custom flash without Triton block pointers

[warrendeng]

The optional segment-anything-fast flash_4 kernel still uses Triton's legacy tl.make_block_ptr and tl.advance APIs, which were removed by triton-lang/triton#10833. This causes the full sam_fast TorchBench model to fail during eager validation with newer Triton revisions.

Detect those APIs before importing segment_anything_fast and disable the kernel with SEGMENT_ANYTHING_FAST_USE_FLASH_4=0 when they are unavailable. On Triton 3.9+, the benchmark continues exercising the full model through its existing SDPA fallback, while older Triton behavior remains unchanged.

Test plan:

• python3.12 -m py_compile torchbenchmark/models/sam_fast/__init__.py
• Verified with Triton 3.7 that the custom kernel remains enabled when both APIs exist
• Simulated missing tl.advance and verified the adapter disables the custom kernel before package import

[warrendeng]

Both CI failures exist on main:

• build-push-docker: nightly torchvision requires torch==2.12.0.dev20260407, which is unavailable. Same failure on main
  (https://github.com/pytorch/benchmark/actions/runs/32268078372).

• Test cpu: eight Omniglot tests fail loading checkpoints with weights_only=True; NumPy’s _reconstruct is not allowlisted. Same failures on main
  (https://github.com/pytorch/benchmark/actions/runs/32135902249).

[meta-codesync]

@warrendeng has imported this pull request. If you are a Meta employee, you can view this in D116691248.