PyTorch drops obsolete Python wrapper symbol extraction in codegen

A recent PyTorch commit (PR #184460) by contributor jansel eliminates an obsolete step in the Python wrapper code generation: the recovery of symbolic sizes and strides from tensor inputs. In the old approach, the code generator had to inspect tensor inputs to extract these symbolic values, adding unnecessary complexity and potential for errors. The fix ensures that free symbols (symbolic dimensions and strides) are now passed directly as explicit graph inputs to the codegen, streamlining the process.

The change only affects the Python wrapper path; the C++ wrapper remains unchanged because it still manages its own size/stride binding behavior. This cleanup is part of ongoing efforts to modernize PyTorch's internals, making the codebase more maintainable and efficient. The PR also fixes issue #102870, likely related to a bug caused by the outdated extraction method. For PyTorch developers and contributors, this means cleaner codegen logic and a reduced surface for subtle bugs in dynamic shape handling.