perf: vectorize 2D sarcomere tracker via SoA + numba greedy claim

Replaces the per-track QueryPoint dataclass loop with struct-of-arrays
state and a numba-compiled _greedy_claim kernel. Same algorithm and
output, substantially faster on real movies. Adds _as_arr_padded helper
so per-detection attribute arrays always align with the detections array
even when callers pass shorter lists.

Author: danihae