bugfix: bugfix to #949

[yzh119]

The sm86/sm89 version of mla kernel was not tests after change #942, this PR fixes the issue.

This PR also make the following changes:

1. adding the mla unittest to CI (on a10g node).
2. shrinking the unittest of mla so that CI can finish in reasonable time.
3. change is_sm90a_supported(torch.device("cuda")) to backend == "fa3" and not is_sm90a_supported(torch.device("cuda")): for non-hopper GPUs, as pointed by @Atream .

[yzh119]

ci has some issues similar to https://discuss.pytorch.org/t/torch-pytest-leads-to-memory-fragmentation-how-to-do-proper-integration-testing-of-a-lot-of-torch-models/201231

The oom hook (

flashinfer/tests/conftest.py

Lines 124 to 135 in 6d5320b

	@pytest.hookimpl(tryfirst=True)
	def pytest_runtest_call(item):
	# skip OOM error
	try:
	item.runtest()
	except (torch.OutOfMemoryError, RuntimeError) as e:
	if isinstance(e, torch.OutOfMemoryError) or "CUDA error: out of memory" in str(
	e
	):
	pytest.skip("Skipping due to OOM")
	else:
	raise

) didn't totally avoid the issue.

[yzh119]

Temporary solution is to set environment variable PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to reduce fragmentation.