viable/strict/1776438505: Codegen mutation epilogue in _create_runtime_wrapper (#179600)

The PyTorch team at Meta has merged a significant optimization to the framework's compilation pipeline. Pull request #179600, titled "Codegen mutation epilogue in _create_runtime_wrapper," replaces the runtime loop that processes mutated inputs with a straight-line, code-generated function. This function is created at compile time and directly handles operations like `set_`, `as_strided_`, `copy_`, and `detach().copy_` based on the mutation metadata of each input tensor, rather than checking flags during execution.

The change specifically targets how PyTorch's Dynamo compiler handles in-place operations (mutations) on tensors, such as `x.mul_(2)` or `a.add_(1)`. Previously, a Python loop would iterate through `mutated_inp_runtime_indices` at runtime to apply these changes. Now, the compiler generates a custom `_apply_mutations` function that contains only the necessary copy operations for the specific mutation pattern of the compiled graph. For example, a single mutation generates `orig_inputs[0].copy_(updated_inputs[0])`, while a leaf mutation under `no_grad` generates a conditional branch checking `requires_grad`.

Benchmark results show the performance improvement is subtle in absolute terms—speedups of 1.01x to 1.03x—because the actual `copy_` operations dominate the runtime cost. The primary advantage is architectural: it removes Python interpreter overhead from the mutation application path and resolves complex branching logic ahead of time. This makes the runtime execution more predictable and efficient, especially for graphs with many potential mutation paths. The PR also updated tests to use `requires_grad` inputs and removed a metadata-only test that caused graph breaks.