BREAKING CHANGE: v0.16 -> CBQ input/output implementation

[alejandrogzi]

This PR includes the necessary implemenation fordeacon filter to accept CBQ (BINSEQ) input, detected automatically, and writes CBQ when the output path ends in .cbq. Benchmark output is consistent between any format combination.

Main change against #31 is only CBQ support, dropping VBQ/BQ.

Benchmark: FASTQ vs CBQ, deacon filter 0.16.0

Input: 500k real viral reads (~4 kb each, 1.9 GB) from rsviruses17900 (Zenodo), filtered against an index built from the 17,900 virus contigs (rsviruses17900.fa.gz); keep-all thresholds (-a 1 -r 0), -t 8, 16-core box; hyperfine min/median of 3 + one /usr/bin/time -v pass for peak RSS.

arm	min (s)	median (s)	peak RSS (MB)
fq->fq	2.95	3.05	662
fq->fq.gz	4.02	4.06	681
fq->cbq	2.86	2.91	674
fq.gz->fq.gz	4.26	4.32	687
fq.gz->cbq	3.08	3.09	671
cbq->fq	2.98	3.01	813
cbq->fq.gz	4.01	4.03	830
cbq->cbq	2.98	3.06	840
fq(r1+r2)->fq	3.27	3.46	663
fq(r1+r2)->cbq	3.05	3.16	675
cbq->fq (paired)	3.31	3.36	814
cbq->cbq (paired)	2.96	2.98	841

Findings

• CBQ output is free: fq->cbq (2.86 s) beats plain fq->fq (2.95 s) + fq->fq.gz (4.02 s); gz output is the slowest arm by ~40%.
• CBQ beats gzip on size: 197.5 MB vs 330.1 MB (40% smaller) on this data; 10× smaller than plain fq.
• CBQ input costs ~150 MB RSS (mmap + embedded index: 662→813 MB); negligible on speed.
• gz input adds ~0.2 s regardless of output format. CBQ roundtrips preserve counts exactly (seqs_out identical within each input group).

*Paired reads were derived by half-splitting real reads.