fix dihedral_backbone() returning garbage angles across chain breaks

[steps-re]

fixes #744

dihedral_backbone() builds its phi/psi/omega coordinate arrays purely by positional adjacency between residues in the AtomArray. it never checks whether two consecutive residues are actually bonded. so if a structure has a missing loop, or you concatenate two unrelated fragments into one AtomArray, the dihedrals spanning that gap come out as finite garbage instead of NaN, exactly like padix-key called out in the issue thread.

went with the approach discussed in the thread: check the C-N distance between consecutive residues directly on the coordinates dihedral_backbone() already extracted, rather than calling check_backbone_continuity() separately (which would mean re-filtering the whole array again). used the same 1.2-1.8 A tolerance that check_backbone_continuity() / filter_linear_bond_continuity() already use elsewhere in the codebase, so the threshold is consistent across the library. where the C(i)-N(i+1) distance falls outside that range, psi[i], omega[i], and phi[i+1] get set to NaN since those are the three angles that straddle the junction between residue i and i+1.

while testing this i found nucleotide_dihedral_backbone() has the exact same bug for the O3'-P bond between residues, so I fixed that too with the same approach (epsilon[i], zeta[i], alpha[i+1] get NaN'd on a broken O3'(i)-P(i+1) junction). that one isn't filed anywhere, just the same root cause turning up twice.

verification:

• added test_dihedral_backbone_chain_break and test_nucleotide_dihedral_backbone_chain_break to tests/structure/test_geometry.py. both build a synthetic structure from two real fragments of 1l2y / 4p5j, renumber the second fragment's res_id to be perfectly continuous with the first (so the test can't accidentally pass just because of a res_id jump), then translate it 1000 A away so the two fragments are not bonded. confirmed both tests fail on current main (angles come back as finite numbers) and pass with this fix.
• ran the full tests/structure/ suite (4176 passed, 13 skipped, no failures) to check for regressions, plus the existing multi-model, MDTraj-consistency, and barnaba-consistency dihedral tests specifically.
• ruff check + ruff format clean on the touched files.

disclosure: this fix (code + tests) was put together with AI assistance (Claude), then reviewed and verified by me before opening the PR.