Fix infer_cbeta corrupting the cached inferred_cbeta array

[winklemad]

No linked issue — found by reading the code.

The bug

ProteinChain.infer_cbeta() builds a new chain whose glycine CB positions are NaN (unless infer_cbeta_for_glycine=True). To do so it takes the cached inferred_cbeta array by reference and NaN-fills the glycine rows in place:

inferred_cbeta_positions = self.inferred_cbeta       # cached_property -> shared array
if not infer_cbeta_for_glycine:
    inferred_cbeta_positions[np.array(list(self.sequence)) == "G", :] = np.nan

Since inferred_cbeta is a @cached_property, this permanently corrupts the original chain's cached array, and everything derived from it (e.g. pdist_CB, which is squareform(pdist(self.inferred_cbeta))). So the value of inferred_cbeta / pdist_CB silently depends on whether infer_cbeta() was ever called:

chain.inferred_cbeta[gly_idx]   # -> real inferred CB coordinate
chain.infer_cbeta()             # default call; intended to be side-effect-free
chain.inferred_cbeta[gly_idx]   # -> [nan, nan, nan]   (cache mutated)

Why it's a bug (not intended)

• The same method already copies the other arrays before mutating them: atom37_positions = self.atom37_positions.copy() / atom37_mask = self.atom37_mask.copy().
• The sibling ProteinComplex.infer_cbeta computes inferred_cbeta_positions into a fresh local, so it isn't affected.

Fix

Copy the array before the in-place NaN-fill (self.inferred_cbeta.copy()). The returned chain is unchanged (glycine CB still masked by default, kept with infer_cbeta_for_glycine=True); only the mutation of the shared cache is removed.

Testing

Added esm/utils/structure/protein_chain_test.py::test_infer_cbeta_does_not_mutate_cached_inferred_cbeta, asserting that inferred_cbeta and pdist_CB are unchanged after calling infer_cbeta(). It fails before this change and passes after. ruff check / ruff format are clean.