Add finite-replica confidence floor and within-source block reference for multi-replica edge confidence

[MarvinTaterra]

Summary

MDPath already reports a per-edge multi-replica confidence (-confidence). Two things
were missing to make that number interpretable:

1. It has a hard floor that depends on how many replicas you ran. With R
   non-negative replica values the coefficient of variation is bounded, so
   confidence = 1/(1+CV) cannot fall below c_min = 1/(1+CV_max). Without that bound
   a "0.6" from 3 replicas and a "0.6" from 20 replicas look the same but aren't.
2. There was no baseline. A high confidence may simply be what finite sampling
   produces on its own, with no genuine replica-to-replica agreement behind it.

This PR adds both, plus the visualization and tests to go with them. Everything is
opt-in: without the new flags, behaviour is identical to current master.

What's new

Finite-replica floor (mdpath/src/confidence.py)
edge_confidence.csv gains c_min (the attainable lower bound), c_norm (the score
rescaled from [c_min, 1] onto [0, 1]) and ddof.

Within-source block reference (mdpath/src/null_model.py, new)
Each source trajectory is cut into contiguous, non-overlapping blocks of length L;
random groups of R blocks are resampled to estimate how reproducible an edge looks
from within-trajectory sampling alone. Observed replicas are subsampled to the same
(L, R) so the comparison is matched. The headline output is

rho = cv_obs / cv_null

i.e. does the between-replica dispersion exceed the within-source reference?

Null-aware visualization (mdpath/src/path_confidence_viz.py)
The existing ChimeraX viewer's hover gains the reference metrics; colouring, tube
widths, spline geometry and structure embedding are unchanged. An opt-in second viewer
(-experimental-c-corr) colours by the experimental corrected score.

Terminology (deliberate)

• confidence / c_norm are dispersion-based stability scores, not probabilities,
  significance levels, or effect sizes. An edge with consistently negligible MI can
  score highly.
• The block analysis is a within-source sampling-stability reference, not an
  external "no-coupling" null — the blocks are real trajectory windows and still contain
  genuine dynamics. It is named as such throughout.
• c_corr_experimental (quadrature CV subtraction) is an unvalidated heuristic. It is
  opt-in, labelled experimental in the CLI/metadata/legend, and finite only when
  corr_status == "above_reference"; undefined and at-or-below-reference edges stay NaN
  so they are never rendered as perfect confidence. rho is the recommended output.

New CLI flags

Flag	Default	Meaning
-null {none,block}	none	Run the block reference (implies -confidence).
-null-block-sweep	25,50,100,250	Block lengths (ns); infeasible ones dropped with a warning.
-null-block-ns	100	Primary block length for the per-edge table.
-null-sources	all	Which replicas supply reference blocks.
-null-draws	100	Resampling draws.
-null-seed	0	Seed.
-ddof {0,1}	0	Delta-DoF for std/CV and the floor.
-null-outdir	null_analysis	Output directory.
-experimental-c-corr	off	Opt-in experimental c_corr viewer.

mdpath -top system.pdb -traj rep1.dcd rep2.dcd rep3.dcd rep4.dcd \
       -confidence -null block -null-block-sweep 25,50,100,250 -null-block-ns 100