Recover anchors from filtered hits after chaining

[ksahlin]

This is a rework of the anchor recovery in #622 (Nicolas Buchin) with
the same output, about half the code and a faster implementation.

Hits whose hash is too frequent are filtered before chaining. After
chaining, they are now looked up again, restricted to diagonal bands
around the extracted chains, and the best run of their anchors is
inserted into each gap of each chain (and before and after it). As in
#622, chains are extracted down to 0.5 times the best score, at most 200
per orientation (new option --mc).

The output is byte-identical to #622 (rebased on v0.19.0) in
mapping-only and in extension mode on 29 simulated datasets (chrY,
CHM13, maize B73 and Mo17, rye Lo7 and Lo7_V3, fruit fly; 50 and 150 nt,
5, 10 and 20 kb). The implementation differs:

• Recovery is a child module of the chainer (src/chainer/recovery.rs),
  so chain.rs and chainer.rs do not move: 6 files change (+558/-2)
  instead of 15 (+1054/-19).
• Filtered hits that share index entries (a repeat in the read) scan
  them once. A bitmap of the reference bins near the chains rejects most
  entries, and runs of entries before the next window are skipped.
• The gap DP skips predecessors at the same reference position (they
  cannot link) and those whose score cannot be improved, using the score
  without the log2 term as an exact upper bound.
• Buffers are reused between the chains of a read.

Total mapping time relative to v0.19.0, #622 -> this change (Apple M4
Max, -x, 1 thread, median of 3 runs; 5, 10 and 20 kb reads):

• chrY: 2.66-1.49x -> 1.88-1.30x
• CHM13: 3.57-2.62x -> 1.70-1.50x
• maize B73: 3.10-2.29x -> 1.75-1.42x
• maize Mo17: 4.20-3.49x -> 3.01-2.55x
• rye Lo7: 3.93-3.21x -> 1.77-1.58x
• rye Lo7_V3: 8.60-5.84x -> 4.73-4.12x
• fruit fly: 1.18-1.08x -> 1.13-1.06x
• 50 and 150 nt reads: 7.10-2.06x -> 4.11-1.61x

Short reads remain expensive: recovery runs on up to 200 chains per
orientation (about 30 per 150 nt read). That cost comes from the chain
extraction settings, which this change keeps as in #622.