perf(fasta,fastq): high-throughput parsing optimizations

[nh13]

Summary

Optimize FASTA and FASTQ record parsing throughput with zero new dependencies.

FASTQ:

• Add bulk newline scanning to read_record(): find all 4 newlines in a single memchr pass when the record fits in the BufRead buffer, avoiding 4 separate BufRead operations per record. Falls back to the existing line-by-line path for records spanning buffer boundaries.
• Add a Reader Builder with 256 KiB default buffer capacity to maximize the bulk parsing hit rate.

FASTA:

• Reuse sequence buffer across Records iterations via std::mem::replace, eliminating repeated Vec growth from zero for similar-length records.
• Optimize consume_empty_lines() to count contiguous \r/\n bytes in a single fill_buf() call instead of two calls per newline character.

Benchmarks

Real data, from disk:

Test	Before	After	Speedup
FASTQ 176 MB short reads (default 8K buf)	88ms	75ms	1.17x
FASTQ 176 MB short reads (Builder 256K buf)	88ms	63ms	1.40x
FASTA 3.0 GB GRCh38 reference	1.83s	1.55s	1.18x

In-memory throughput (Criterion):

Test	Before	After
FASTQ read_record 150bp × 100K	4.4 GiB/s	6.0 GiB/s
FASTQ Builder 150bp × 100K	—	6.1 GiB/s
FASTA records() 1Kbp × 10K w80	1.6 GiB/s	2.0 GiB/s
FASTA records() single-line 1Kbp × 10K	2.0 GiB/s	6.0 GiB/s

Test plan

• All existing tests pass (cargo test -p noodles-fastq -p noodles-fasta)
• Benchmark suite runs (cargo bench -p noodles-fastq --bench reader / cargo bench -p noodles-fasta --bench reader)
• No public API changes to existing types — Reader::new(), read_record(), and records() work identically
• Bulk path correctly falls back for records spanning buffer boundaries (tested with BufReader::with_capacity(8, ...))
• Edge cases: CRLF, quality containing @/+, empty sequences, missing trailing newline