Dataloader output: dict instead of 4-tuple

[urancon]

What

NeuralDataset.__getitem__ and neural_collate now return a dict
({'stims', 'responses', 'valid_mask', 'stim_meta'}) instead of a positional
4-tuple.

Why

So future per-trial variables (behaviour, pupil size, running speed — as in the
Sinz-lab neuralpredictors / Sensorium datasets) can be added as extra keys
without changing the unpacking contract. neural_collate passes any unknown
per-item key straight through as a length-B list, so a dataset only has to add
the key to its __getitem__ dict:

def __getitem__(self, i):
    return {'stims': ..., 'responses': ..., 'valid_mask': ..., 'stim_meta': ...,
            'behav': self.behavior[i]}   # -> batch['behav'] for free

This is a hard break to positional unpacking, done pre-0.1.0 while there are no
published users.

Changed

• Core: __getitem__ + neural_collate (library); Fitter (both loops).
• Tests: all consumers + the shared _ToyDataset.__getitem__.
• Docs: README quickstart, fitter.md, data_paradigm.md,
  metrics_paradigm.md; full rewrite of the stale README_datasets.md
  conventions page (it documented the long-gone (spectrogram, responses, ccmax, ttrc) tuple).
• Notebooks: the 4 that unpack a batch manually (crcns_aa_tutorial,
  dataset_concatenation, explore_nat4, alice_eeg_tutorial). Source updated;
  a re-exec pass is folded into the pre-release notebook smoke test.

Verification

• pytest -m "not slow" -> 440 passed (~17 s)
• ns1 / downer / wingert slow integration files -> 87 passed
• full suite -> 575 passed, 18 skipped (identical to the pre-refactor
  baseline — zero regressions)

🤖 Generated with Claude Code