[externalization] Add a two-step patch/sub-export API for module externalization - #53

Cross-reference the two "no call sites" warning paths so neither reads as
dead code left over from the refactor: _drop_missing_call_sites handles
top-level marked submodules missing from the whole-model program, while
_PreparedModules.__iter__ handles the nested case it cannot see, once the
enclosing module's sub-export exists. Both docstrings now point at each
other, and _subexport_and_restore names the split.

Add test_externalize_multiple_staged_entries_numerics: two staged entries
in one to_coreai() call, one via add_pytorch_module(externalize_modules=)
and one plain add_exported_program, validating numerics on both
entrypoints. This pins the invariant that externalization is per-entry —
_init_conversion_state() resets _externalized_exported_programs before
each entry, so a plain entry converted alongside an externalized one
stays flat. Verified non-vacuous: dropping that reset and adding a third
_perform_externalization call site fails the test on a duplicate
rms_norm symbol.

_validate_numerics takes a function_name argument so it can check a
named entrypoint instead of only "main".

The module docstring documents four Phases (Mark, Prepare, Export, Emit)
while individual docstrings still used a separate six-Step scheme that no
longer lined up — export was "Step 4" but Phase 3, prepare was "Step 3"
but Phase 2, and there was no Step 2 at all. Relabel every per-function
reference to the Phase vocabulary and drop the number from
_restore_externalized, since restore is no longer its own phase.

Docstrings only; no behavior change.