test: restore sequence diversity in prefilter fixtures

[mikessh]

Closes #103. Fix authored by @marcus-campbell — cherry-picked from
https://github.com/marcus-campbell/arda/tree/fix-prefilter-random-fixtures with authorship preserved.

"".join(random.Random(seed).choice("ACGT") for _ in range(n)) builds a fresh RNG per draw, so
every choice() returns the same base. Six prefilter fixtures were homopolymers, not varied
deterministic DNA — seeds 1/3/8 gave poly-C, 2/6 poly-A, 5 poly-G. The tests passed while
exercising almost none of the diversity they read as covering: test_unrelated_sequence_is_rejected
was asserting k-mer specificity against a poly-C target.

The fix hoists the RNG into _random_dna(seed, length) and advances it once per base.

test_keep_records_indexes_the_fasta_mmseqs_searches gets a second fix. Its FASTA is deliberately
line-wrapped at 150 to assert _read_fasta_seqs joins the lines, but the query target[100:200]
would still hit on k-mers wholly inside line 1 — so a reader that indexed the two lines separately
passed anyway. The query is now target[142:158]: the single 16-mer straddling the wrap, present
exactly once in the joined sequence and in neither line alone.

No source change; tests only.

Verification

• tests/unit/test_prefilter.py — 10 passed
• tests/unit tests/synthetic tests/realworld — 968 passed, 9 skipped
• ruff 0.15.9 check src/ tests/unit/test_prefilter.py — clean

🤖 Generated with Claude Code