[inductor] FX graph cache: Fix bug handling constants (#121925)
Summary: During key calculation for FX graph caching: Rather than specialize on "small" vs. "large" tensor constants (i.e., inlined vs. not inlined), always hash on the tensor value. Doing so avoids the complication of trying to later attach the constant values as attributes to an already-compiled module. Instead, different constants will cause an FX graph cache miss and we'll just compile.
Test Plan: New unit test
Pull Request resolved: https://github.com/pytorch/pytorch/pull/121925
Approved by: https://github.com/eellison