[FIX] Skip the profiler buffer load check when Triton has no active driver

[mark14wu]

Summary

On CPU-only hosts, every launch traced with the Profiler failed before the kernel ran:

RuntimeError: 0 active drivers ([]). There should only be one.

The buffer load check is enabled by default (PROFILER_DISABLE_BUFFER_LOAD_CHECK=0), so Profiler.pre_warmup_callback asks TritonTrace.run to call the real jit_fn.warmup before each launch. Warmup compiles the kernel through triton.runtime.driver.active, which raises when no backend driver is available. The Profiler is the only client that requests warmup. The existing profiler tests all set profiler_disable_buffer_load_check = True, so CPU CI never ran this path.

Changes

• Profiler.pre_warmup_callback checks driver.active first. With no driver, it skips warmup and turns off the buffer load check for that profiler, so the launch runs normally.
• The finalize report now says Skipped: no active Triton driver to compile the kernel (e.g. CPU-only host). instead of dropping the section or flagging a buffer load issue without any ASM to check.
• GPU hosts are unchanged: warmup still runs and the check reports the same as before.

Tests

• Added test_buffer_load_check_without_active_driver in tests/end_to_end/test_profiler.py. It monkeypatches Triton's driver factory to simulate a driverless host, so it also covers this path on GPU machines. Without the fix it fails with the same 0 active drivers error.
• pytest tests/end_to_end/test_profiler.py tests/unit/test_profiler.py: 24 passed.
• Manual repro with CUDA_VISIBLE_DEVICES="" and a masked add kernel under @tilelens.trace("profiler"): it used to fail with 0 active drivers and now completes with the skip note. With the GPU visible (RTX 4090), the output is the same as before.
• pre-commit (ruff, ruff-format, mypy, codespell, …) passes on the changed files.

[github-actions]

Performance Benchmark

Benchmark	main (min)	PR (min)	Change	Samples
gemm	0.104s	0.103s	-0.9%	20 / 20
gemm_oob	0.116s	0.116s	-0.0%	20 / 20
indirect_load	0.021s	0.021s	+0.4%	20 / 20
nested_loop	0.233s	0.232s	-0.1%	20 / 20
block_pointer_loop_advance	0.124s	0.123s	-0.2%	20 / 20
liger_jsd	0.138s	0.138s	+0.3%	20 / 20
flaggems_layernorm	0.392s	0.391s	-0.3%	20 / 20
swiglu	0.169s	0.169s	+0.0%	20 / 20
cross_entropy	0.965s	0.967s	+0.1%	20 / 20
fused_linear_jsd	0.209s	0.208s	-0.2%	20 / 20
Total	2.470s	2.470s	-0.0%	N/A

Iterations: 1 warmup + 20 measured
Samples are shown as main / PR; long pytest benchmarks may use fewer samples.