Let skew take a directory of FASTQs

[micah-olivas]

What

usortm skew now accepts a directory of FASTQs, not just a single file.

The reads argument resolves three ways: a single FASTQ, a directory searched
recursively for .fastq/.fq files (gzipped or not), or a mix of both. Files
found in a directory are counted together as one library, so a basecaller's
fastq_pass/ tree can be passed as-is with no concatenation step.

How

• collect_fastqs() in usortm/qc/counting.py does the resolution. Non-FASTQ
  files in the tree are ignored; a directory holding no reads raises rather
  than silently counting nothing.
• count_variant_reads() passes every file to a single minimap2 invocation.
  minimap2 accepts any number of query files and emits each read's alignments
  consecutively, in input order, so the per-read grouping logic is unchanged
  and no temporary concatenated copy of the reads is written.
• count_fastq_reads() sums across files for the progress total.
• The file list is recorded in skew_report.json under fastq_files, and in
  the project file under measured_skew. A directory listing can change
  between runs, so the report states which files were actually counted.

Tests

Full suite passes (262). Seven new tests cover resolution of files,
directories, mixed input, and the empty-directory error. The load-bearing one
splits a library FASTQ into a plain half and a gzipped half in a subdirectory
and asserts the per-variant counts match the single-file run exactly, so
recursion, gzip handling, and pooling are covered end to end.

Note on the first commit

This branch carries two commits. 9d5929e is a snapshot of the skew feature
that was still uncommitted in the working tree when this change was written;
it is the base so that 588c41a shows the directory-input change on its own.