[Feat] Add FoldPythonScalarConstantsPass to fold pure Python scalar nodes before piecewise split

[Young-Leo]

What

Add FoldPythonScalarConstantsPass to the full-graph pass pipeline (runs before split_module). It evaluates an allowlist of pure scalar operators whose inputs are already plain Python scalars, replaces the FX node with the resulting literal, and removes the now-dead node.

Why

After a TensorifyScalarRestartAnalysis recapture, a tensor-derived SymFloat (e.g. the FP8 clamp bound -448.0) can come back as a concrete Python constant while the scalar arithmetic stays as an FX node (operator.neg(448.0)). When that value is reused across a piecewise boundary, split_module promotes the node to a subgraph output — and Inductor rejects a bare Python float as a compiled-graph output, crashing compilation.

torch has no native pass that folds Python-scalar FX nodes into literals.

• torch._inductor.constant_folding only folds tensor constants (it explicitly skips nodes with no tensor input, and materializes results as get_attr buffers), and it runs inside each subgraph after split_module — too late for our case.
• torch.fx.experimental.const_fold.split_const_subgraphs restructures the graph via its own split_module, conflicting with our piecewise splitting. So this small, pre-split scalar-literal fold fills a real gap rather than duplicating existing functionality.

Scope / Safety

• Only an allowlist of pure scalar operators is evaluated.
• Folds only when all inputs are already plain Python scalars; Tensor / SymInt / SymFloat / runtime-dependent values are never touched.
• Arithmetic errors are caught and skipped; runs unconditionally on every (re)capture, so it also handles graphs that no longer contain item().

Tests

• Unit tests: constant-chain folding, no-op on tensor/symbolic inputs, and no execution of non-allowlisted call_functions.
• End-to-end regression reproducing the FP8 clamp-across-boundary crash.

[jiahy0825]

LGTM