Seeding for build_vset, two rounds of bug fixes, Python 3.13, new logo - #60

Seeding: build_vset(..., reps=n, seed=s) gives each vfunc its own
random_state drawn from np.random.SeedSequence(s), so repeated
perturbations (e.g. bootstrap samples) are reproducible.

Bug fixes (each has a regression test in tests/test_bugfixes.py):
- filter_vset_by_metric kept nothing for build_vset output without
  param_dict or with reps, double-counted configurations across
  upstream perturbations, and could not filter a filtered Vset
- dict_to_df(param_key=...) mixed up parameters across rows and split
  values containing '='
- perturbation_stats: operator-precedence bug, string/callable func,
  prefix applied to default columns
- PCSPipeline.run did not chain steps; generate_names crashed
- MLflow tracking: params only logged in some cases, runs never closed,
  non-numeric outputs (fit/predict) crashed logging
- caching served stale results after a function was redefined and
  failed on lambdas; cache now keyed on a fingerprint of the inputs
- Vset accepted bad containers and duplicate keys, ignored async
  vfuncs given as a dict, and dict keys were not turned into Subkeys
- predict/predict_proba on a Vset without a predictor now raise
- lazy async outputs feeding a lazy sync Vset were not resolved
- Subkey hash disagreed with equality
- cum_acc_by_uncertainty gave an unclear error on class predictions

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Xy8Dg7fAqocHGzMqVCiSK

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Xy8Dg7fAqocHGzMqVCiSK

- caching: key the cache on a by-value hash of the inputs (joblib.hash,
  without the lineage entry) so equal copies, e.g. outputs loaded back
  from the cache, hit it; add the source of module-level functions so
  editing one does not serve stale results
- MLflow: metric names from build_vset tuple keys are sanitized; runs
  are closed even if logging fails; runs are keyed by full subkeys so
  equal values from different Vsets are not merged
- PCSPipeline.run passed Vfunc objects on from an already-fitted step
- vfunc_keys that compare equal (1 and 1.0) were not caught as duplicates
- perturbation_stats on array entries: tuple group keys (build_vset
  models, predict_proba with uncertainty grouped by model) crashed,
  split=True crashed on 2-D entries (now out<i>_<j> columns), and
  unsupported funcs were silently ignored (now raise)
- build_vset: partials of one function get distinct names, and distinct
  functions or values that share a key raise instead of being dropped

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Xy8Dg7fAqocHGzMqVCiSK

The docs header uses the wordmark, the favicon is the drop ring, and the
README points at the new wordmark. logo_vflow_straight.png is kept since
external pages may link to it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Xy8Dg7fAqocHGzMqVCiSK

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013Xy8Dg7fAqocHGzMqVCiSK