PyTorch Inductor update lets standalone_compile reuse caller FakeTensorMode

PyTorch’s Inductor JIT compiler has received a targeted optimization for its `standalone_compile` API. Previously, calling `standalone_compile(..., dynamic_shapes="from_example_inputs")` always constructed a fresh `FakeTensorMode` with its own `ShapeEnv`. This meant that even if a caller had already created fake tensors under a custom mode, the standalone tracing context would ignore it and allocate a completely new mode, wasting resources and preventing the caller from sharing shape constraints.

To address this, the latest PR (merged by jansel) introduces an optional `fake_mode` parameter to `standalone_compile`. When passed, the standalone context uses that mode instead of generating a new one. The parameter is only honored with the `from_example_inputs` dynamic shapes strategy; other strategies derive their mode from the tracing context or graph metadata, so passing `fake_mode` there is rejected to avoid silent misbehavior. This narrow, well-defined change lets advanced users pre‑create fake tensors on the host and reuse them during compilation, eliminating redundant device allocation. The PR includes thorough test coverage to ensure the new argument integrates cleanly with existing cache and dynamic shape tests. For engineers working with PyTorch’s graph compilation pipeline, this update reduces overhead and makes the Inductor API more composable.