perf(classify): pack move tables; vote once per read

[jayhesselberth]

Closes #479

(1) Packed move table: shipped (bar cleared)

AnchoredRead.seq_to_sig: Vec<i64> (8 B/base, held for the whole BAM) becomes moves: PackedMoves (one bit per stride) with AnchoredRead::seq_to_sig() deriving the map where used. A unit test pins the packed expansion equal to the old seq_to_sig_map over 200 random move arrays (values 0/1/2, trailing zeros, non-byte lengths).

Measured on ldx14_frozen_edx06 (70,188 anchored reads, 65,821 classified, 2 POD5, --threads 8, profiling profile), anon RSS peak from /proc/PID/smaps_rollup (Anonymous:, polled 0.2 s), arms interleaved, 3 reps on one node:

arm	peak anon RSS (KB)
main	394292 / 393880 / 398492
this PR	268360 / 275616 / 268468

About 125 MB saved, 32% of peak (bar: 10% at test scale or 1 GB at 57 M reads), ~1.8 KB/read, so ~100 GB extrapolated at 57 M reads. Wall unchanged within noise (9.4-10.5 s warm; the first main rep was cold, 38 s).

Identity: cl tags and TSV byte-for-byte equal after sorting in all 3 reps (BAM record bodies equal in order, 73,551 records; TSV 70,189 lines).

(2) One vote per read: shipped (correctness fix)

Votes are now tallied from the deduped map after the scan (BamScan::tally_votes), not per record. a_read_with_several_alignments_votes_once_on_the_one_kept fails with the old per-record counting (time: 3, reversed: 1, total 4 instead of 2). On the real run the frame vote is unchanged (time=0, reversed=66943 on both): secondary and supplementary records are already skipped, so multi-anchored reads do not occur in this set.

Not changed

Scorers, GPU, the waveform path's layout. Library callers: AnchoredRead::seq_to_sig is a method now; the field is moves. Benches updated.

CI-equivalent run clean (fmt, workspace clippy --all-targets, gpu x3, waveform-onnx, models-download, train, crf-decode, classify benches; nextest workspace/cli experimental/classify waveform-onnx; doctests).