fix: keep group index and result position apart in grouped estimation (#318)

[MohammadYusif]

Summary

• EventModel._estim_probs_groups collects the per-group results in a list built positionally
  over np.unique(groups), but then indexes that list with the group itself. The two only
  coincide when every declared group occurs in the data — as soon as a group declared in
  channel_map/time_map has no trial left (a level absent after trial rejection, a
  participant subset missing a condition), event probabilities are written onto the wrong
  trials and the loop ends on IndexError: list index out of range.
• The per-group log-likelihoods had the same problem from the other direction: they were
  returned as a compacted array over only the groups present in the data, while
  group_labels covers every declared group, so group_lkh — and the traces_group /
  xrtraces built from it in EM — silently stopped lining up with the groups they describe.
• This keeps the position in the result list and the group index separate, and sizes the
  likelihood array on the declared groups, leaving nan for a group without trials.

As mentioned in the issue, the alternative would be to reject a declared group without
trials outright. I went with indexing by position because a model fitted on the full data
and then transformed on a subset that happens to miss a condition is a legitimate case, but
happy to switch if you'd rather not allow empty groups at all.

Changes

• hmp/models/event.py: in _estim_probs_groups, iterate with
  enumerate(data_groups) so the result position and the group index stay apart;
  pre-size likelihood on the declared groups (np.zeros(self.channel_map.shape[0]) * np.nan)
  and fill it by group, so group_lkh stays aligned with group_labels; sum it with
  np.nansum, which is unchanged whenever every group is present.
• tests/test_fixed.py: test_grouping_absent_group fits the same trials twice with fully
  independent per-group parameters — once coded 0/2 with the declared group 1 left
  without trials, once coded 0/1 with only the two occurring groups declared — and
  asserts the likelihood and both parameter arrays match, that the group likelihoods are
  reported for every declared group, and that the empty one is nan.

Fixes #318

[GWeindel]

Hi @MohammadYusif

Thank you for your pull request, could we maybe have a quick online meeting regarding your suggestion? You can send me an email at gabriel.weindel@unil.ch