[Attn] Remove erroneous @torch.compile on ParallelAttentionFunction class

[arbi-dev]

Summary

fla/ops/attn/parallel.py applies @torch.compile directly to the torch.autograd.Function class ParallelAttentionFunction. torch.compile is meant for functions / nn.Modules / callables, not for an autograd.Function subclass — the decorated object's innermost fn resolves to the class, which has no __code__.

It has been silently harmless because the decorator is inert for the .apply path that parallel_attn uses. But it hard-crashes import fla the moment torch's caching_precompile is enabled (a torch 2.9+ feature that persists the Dynamo trace cross-process), because caching_precompile eagerly resolves a precompile cache key at decoration time:

fla/ops/attn/parallel.py:719  @torch.compile
torch/_dynamo/eval_frame.py:690  result = DynamoCache.load(fn)
torch/_dynamo/package.py:914     key = CompilePackage.source_id_from_fn(fn)
torch/_dynamo/package.py:684     sha256_hash.update(str(innermost_fn_.__code__.co_firstlineno).encode())
AttributeError: type object 'ParallelAttentionFunction' has no attribute '__code__'

Since it fails at import, a downstream project that enables caching_precompile cannot import fla at all. This removes the stray decorator. The real compute lives in the parallel_attn_fwd / parallel_attn_bwd Triton kernels invoked inside forward/backward, so removing the class-level decorator leaves forward/backward numerics unchanged.

Closes #979.

Test plan

• import fla.ops.attn.parallel with torch._dynamo.config.caching_precompile = True no longer raises AttributeError (was a hard crash before; succeeds after).
• Syntax/AST parse verified.
• I have not run the full GPU op test suite (tests/ops/test_parallel_attn*) for parallel_attn — I don't have that op exercised in my environment, and per the project's "match the reference numerically (fwd + bwd)" principle I'd rather flag that than assert it. Requesting a maintainer run of the parallel-attn fwd/bwd numerics + the standard CI to confirm before merge. Marked draft for that reason.

Breaking changes

None expected. The decorator was a no-op for the .apply code path (autograd handles .apply specially), so output and gradients are unchanged. If the original intent was to compile the Python glue, the correct placement would be @torch.compile on the parallel_attn function (line 771) instead — happy to switch to that if preferred.