[6/6] Add C. elegans worked example; rewrite README around new knobs

[alexanderbates]

Add C. elegans worked example; rewrite README around the new knobs

examples/celegans_worked_example.py is a self-contained end-to-end run
on the bundled C. elegans connectome. Per-seed influence from every
sensory neuron (83 cells -> 46 cell classes after collapsing bilateral
pairs) onto every non-sensory target (187 -> 136 classes), summed per
(target_class, seed_class), log-adjusted via adjust_influence with an
auto-calibrated const = -log(min_nonzero |sum|), and rendered as two
heatmaps grouped by anatomical body_part with average-linkage
clustering within each group. The script encodes its design choices as
constants at the top: LAMBDA_MAX=0.5 (the small graph needs the leading
mode damped or it dominates every column), COUNT_THRESH=0,
EXCLUDE_BODY_PARTS={'pharynx'}, plus C. elegans neurotransmitter
conventions (only ACh and GABA have unambiguous signs; glutamate /
dopamine / serotonin / octopamine are excluded).

The cell-class regex collapses trailing L/R suffixes only -- omitting
the seemingly more "complete" DL|DR|VL|VR alternatives because Python's
leftmost-first alternation would otherwise match DL at position 2 of
AVDL (with lookbehind V) and produce AV instead of AVD.

Two output heatmaps land in docs/images/ alongside four explanatory
images for the README: a source-to-targets schematic and an
adjusted_influence-vs-traversal-depth scatter pulled from the
natverse/influencer R sibling package, and a bespoke annotated linear
dynamical model + a 12s propagation-to-steady-state animation on a
28-node toy graph.

README is restructured around the parameters a user actually tunes:

• Description: derives the rescaling W_tilde = (lambda / lambda_max(W))
  W with the reverb-knob metaphor, names lambda_max as the constructor
  argument, and gives species-specific guidance (~0.99 for whole-CNS
  Drosophila BANC, ~0.5 for C. elegans). syn_weight_measure documented
  as a deliberate 'count' default with the rationale that signed mode
  pairs cleanly with raw counts but distorts 'norm' (columns no longer
  sum to 1).
• Usage: lists all six entry points (DataFrame init plus from_sql,
  from_csv, from_parquet, from_feather, from_numpy), names the required
  edge list and meta columns, and explicitly states that missing
  columns raise an actionable ValueError.
• adjust_influence section: explains the log-+-const transform and the
  exp(-const) junk-node floor, gives the auto-calibration recipe for
  non-BANC datasets, summarises the three output columns and when to
  use which, and notes that calculate_influence runs adjust_influence
  by default so the standalone import is only needed for advanced
  multi-seed aggregation.
• Worked example section: minimal end-to-end snippet, knobs table
  cross-referencing the Description and adjust_influence sections, the
  two heatmaps, and a short data-source attribution to the OpenWorm
  project + White 1986 + Cook 2019.
• BANC dataset section: alongside the Dataverse DOI, lists the Lee
  Lab's public GCS bucket for the Feather edge list, with from_feather
  consumption.
• Citation section between the worked example and Contributing: cites
  Bates, Phelps, Kim, Yang et al. (2025) (bioRxiv 2025.07.31.667571,
  PMID 40766407) in prose and BibTeX, and points users at the Zenodo
  DOI badge for the software citation.
• One-line cross-link at the top of Description to natverse/influencer
  (the R sibling package that wraps this library and provides a native
  R backend).

Removes two pre-existing duplications in the Description (the
"this code computes the influence scores..." sentence appeared twice;
the trailing paragraph on creating a pandas dataframe and silencing
neurons was already covered by Usage).