evo2 dashboard: split generator into atlas (cached acts) + examples (small pass)

atlas: stats from a RANDOM SAMPLE of an extract.py activation store run through the SAE
(no 7B/megatron) + UMAP from the decoder -> features_atlas + feature_metadata. This is the
same store the SAE trained on; if absent, errors pointing at extract.py.

examples: small --examples-fasta through the full engine -> feature_examples. Fixes the
pass-2 OOM (extract per-feature columns + window to --max-example-bp instead of holding
[S, n_features] per sequence) and skips dead features.

Resolves the earlier design: the atlas no longer re-runs the expensive 7B over the whole
corpus. Labels still join from --feature-annotations (#1630).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Author: polinabinder1