feat(icons): align the ICONS selector with the reference implementation

[tongruiliu]

The official code (github.com/princetonvisualai/icons) is now public, so close the gaps that keep this selector from reproducing it:

• pool_groups: the reference gives the 13 MME subtask validation sets a single vote by max-pooling their per-slice influence vectors (icons/generalist.py). Slices can now be grouped the same way instead of each voting separately.
• warn when only one task votes: the consensus then degenerates to plain top-p influence selection, which is easy to miss with task_boundaries unset.
• vote_percentile: pin the paper's selection ratio p for the per-task threshold instead of always deriving it from the schedule (num_samples / len(dataset)).
• clean stale gradient chunks before a collection and fail when the merged chunks do not cover every sample, rather than silently scoring zero rows for missing ones; the old "resuming" path never skipped a sample.
• components.yaml: proj_dim 5120 / seed 0, matching the reference dims.

Voting was checked vote-for-vote against a transcription of icons/generalist.py (percentile threshold, >= comparison, vote sum, top-k) including the pooled and single-task cases.