trunk/164855bb6a8e8708551338a9ee5c2fa23f2709b0: Fix AOT FXIR parallel Triton kernel reload (#185134)

AOT FXIR imports generated Triton kernels through the wrapper code and then stores the JITFunction in the Triton HOP side table. With parallel Triton compile enabled, the worker returns a pickled CachingAutotuner whose JITFunction has had its underlying Python function stripped by prepare_for_pickle(). FXIR then registered that stripped JITFunction, so a fresh runtime compile could reach Triton's source-location lookup and fail with AttributeError because fn.fn was None.

Reload the parent-side JITFunction after resolving the async compile future and before FXIR wraps it for the side table. This keeps the root cause local to the FXIR import path that consumes worker-returned autotuners. The alternative of forcing compile_threads=1 avoids the race but leaves parallel compile broken, so the existing test workaround is not relied on.

Add a regression test that forces subprocess parallel compile with a fresh cache, checks the side-table kernel has a live function, and executes the generated FXIR graph.

Fixes

Approved by: https://github.com/angelayi