Fix circular-coordinate handling and histogram support in NMI

[MarvinTaterra]

Summary

Two independent defects in the mutual-information pipeline, and they amplify each other.
Both affect NMI and therefore everything downstream: graph weights, path weights,
per-replica confidence, the block reference added in #112, and bootstrap path-confidence
intervals.

Defect A — the periodic coordinate was differenced without wrapping

MDAnalysis reports phi on [-180, 180). A residue oscillating near the branch cut flips
between +179 and -179, so np.diff turned a real 2 degree step into -358.

Defect B — the histogram support was derived from each input's own extrema

Within a single analysis that is defensible adaptive binning. It becomes a defect the moment
NMI values from different inputs are compared — which is exactly what the multi-replica
confidence path, the block reference and the bootstrap all do. Two replicas of identical
dynamics, one containing a single larger excursion, get different bin widths, and that
difference flows straight into MI_std -> MI_cv -> confidence.

Changes

• Wrap the dihedral difference onto [-180, 180) in structure.py — the only np.diff
  in the package. Unconditional: this corrects a defect, not a configurable policy.
• hist_range on NMICalculator, default (-180.0, 180.0), threaded through
  EdgeConfidenceCalculator, BlockSplitNull and BootstrapAnalysis so a whole run shares
  one binning. Added as the last parameter in each signature, so no positional call site
  breaks.
• -hist-range LOW HIGH (or -hist-range none) to set the support per run.
• -diagnose-bins and mdpath.src.mutual_information.binning_diagnostic(): per-residue
  histogram occupancy under the chosen support, plus branch-cut exposure (frames with
  |dphi| > 180 before wrapping), written to binning_diagnostic.csv.
• Per-residue exposure is recorded in DataFrame.attrs["branch_cut_crossings"] and preserved
  across the multi-trajectory column alignment (a bare df[common_cols] dropped it).
• Docstring notes in confidence.py and null_model.py that per-replica and per-block values
  are only comparable under a fixed support.

Backwards compatibility

hist_range=None restores the previous behaviour and is an exact no-op — it reproduces
pre-fix NMI values bit-for-bit, asserted in test_hist_range_none_is_exact_no_op. Use it to
reproduce published numbers.

Note the asymmetry: None restores only the binning. The dihedral wrap is unconditional,
because an unwrapped periodic difference is simply wrong.

No CLI, output-file or column-name changes beyond the two new optional flags.

Two exact-value assertions in test_mutual_information.py were recomputed under the new
default (largest change -3.3 % in the plain case; the invert case moves more because
rebinning changed which pair is the maximum). No assertion was weakened, and the pre-fix
values are retained in the backwards-compatibility test.

Tests

New regression tests cover:

1. wrapping across the branch cut (a 2 degree step stays 2 degrees, and |dphi| <= 180);
2. identical bin edges for two inputs with deliberately different extrema;
3. exact reproduction of pre-fix values via hist_range=None;
4. a single contaminated frame no longer changing a clean residue's occupied-bin count.