Look up fewer filtered hits during anchor recovery

[ksahlin]

In satellite arrays, filtered hits occur every few bases along the read,
and looking up each of them again dominates the cost of anchor recovery.
Now at most one filtered hit per read_len / 400 bases of the read is
looked up (25 bases for 10 kb reads). Reads shorter than 400 bases are
not affected: their output is unchanged.

Total mapping time relative to v0.19.0, before -> after (Apple M4 Max,
-x, 1 thread, median of 3 runs; 5, 10 and 20 kb reads):

• chrY: 1.88-1.30x -> 1.52-1.09x
• CHM13: 1.70-1.50x -> 1.37-1.13x
• maize B73: 1.75-1.42x -> 1.42-1.14x
• maize Mo17: 3.01-2.55x -> 1.66-1.27x
• rye Lo7: 1.77-1.58x -> 1.46-1.17x
• rye Lo7_V3: 4.73-4.12x -> 2.69-1.50x
• fruit fly: 1.13-1.06x -> 1.10-1.04x

Accuracy (% of reads mapped within 200 bp of the true position), change
compared with before, on the same 21 datasets:

• mapping-only: +1.66 to +2.04 on chrY, +0.13 to +0.24 on CHM13, -0.30
  to +0.16 elsewhere (largest loss: Mo17 20 kb, -0.30)
• extension: -0.15 to +0.23 (largest loss: Mo17 20 kb, -0.15)

Is-new-baseline: yes