Inductor proposal would let FX wrappers call non-Triton fused kernels

PyTorch's Inductor compiler can emit an FX graph as a callable when the fx_wrapper option is on, but that path only works for Triton kernels. Any other fused kernel, including the C++ fusion Inductor generates for CPU, raises an error instead of producing a graph.

The gap is representational. Triton launches already have a high-order operator, triton_kernel_wrapper_mutation, that can sit inside an FX graph. CPU fusion bindings are ordinary callables: they take a flat list of tensors and write their outputs in place, so there is no operator through which the FX converter can launch them.

aorenste proposed a non-Triton counterpart, compiled_kernel_wrapper_mutation, plus a side table that holds the compiled callables. The operator would take a kernel index into that table, the indices of arguments mutated in place, and the real inputs. That would let the same FX wrapper path cover CPU fusion and other generated kernels, not only Triton.