v0.7.0

Minor: this cycle makes decode parallel-first — the archive is an analysis
substrate, not just cold storage. Additive Rust and Python APIs (stream
selection, a parallel record visitor, FASTA output) plus one
writer-side compatibility note: long-read archives written at the default
level now use chunked quality coding (context modes 5/6), which v0.6.x readers
refuse with an "upgrade fqxv" error. The on-disk format stays 1.0 — this
release reads every archive any earlier release wrote, and short-read archives
are byte-identical in both directions.

Added

• Stream-selective decode (#269): StreamSelection plus
  decompress_records_select / RecordReader::with_selection skip deselected
  streams without entropy-decoding them. The range coders are ~96–98% of
  decode compute, so sequence-only decode of an ONT archive runs ~12× faster.
  decompress_fasta and fqxv decompress --fasta emit names+sequence as
  single-line FASTA, skipping quality entirely. No format change; a full
  selection keeps the canonical full-decode path bit-for-bit.
• Parallel record visitor (#271, #276): decompress_records_par — a
  map-reduce visitor (init/visit/merge) that scales record consumers with
  the block parallelism instead of funneling through one serial channel — and
  decompress_records_par_select, its composition with stream selection
  (parallel sequence-only decode measured 13× on ONT).
• Python: streams= and fasta= (#274): fqxv.open(..., streams="seq")
  decodes only the selected streams; decompress_to_path /
  decompress_to_bytes accept fasta=True; fqxv.remote passes both through.
• Decode thread-scaling benchmark (#270): bench/scripts/decode_scaling.sh
  plus docs/decode-scaling.md measure full-decode throughput versus thread
  count against gzip/pigz/zstd, with byte-identity asserted for every cell.
• Params::block_seq_bytes (#275): the long-read per-block raw-sequence
  budget is a library parameter (0 = platform auto) instead of a hard-wired
  constant.

Changed

• Long-read quality is coded in parallel-decodable chunks by default
  (#279; design and measurements in #277/#278). Quality — 92–98% of long-read
  decode — was one adaptive coding pass per block; it is now a serially-coded
  warmup prefix plus K−1 whole-read chunks that encode and decode in parallel
  (context modes 5/6, K = 8, Nanopore + PacBio only). Measured on ONT/HiFi:
  full decode ~1.9× faster at 16–32 threads (ONT 17.7 → 9.1 s, HiFi 66.5 →
  33.8 s), HiFi compress 17% faster, for +0.36% / +0.19% archive size.
  Compatibility: v0.6.x binaries cannot read mode-5/6 archives — they
  refuse loudly (unsupported quality context mode 5; this archive needs a newer fqxv), never misread. Short-read archives are byte-identical, and
  --max (level 9) pins the serial smallest-archive layout, also
  byte-identical — chunking is a default-level trade only, and
  Params::quality_chunks (0 auto / 1 serial / K ≥ 2) opts out or tunes it.
• Nanopore default block budget dropped 256 → 64 MiB of raw sequence
  (#275): the benchmark ONT file cuts 5 blocks instead of 2, unlocking
  block-parallel decode (measured 3.2× at 8 threads) for +1.1% archive size.
  --max and explicit --block-reads keep maximal blocks. Combined with
  chunked quality, the flat ~11 MB/s ONT decode becomes 8.3 → 67 MB/s over
  1→32 threads (docs/decode-scaling.md).
• Bumped the cargo minor/patch dependency group (#268).