refactor: drop Transcript::read_seq, a public field nothing reads

[BenjaminDEMAILLE]

Transcript::read_seq is filled with a full copy of the read at every finalised alignment, and nothing in the crate reads it.

How I know nothing reads it

Not from grepping. I deleted the field and asked the compiler. All 97 resulting errors are struct literals or assignments:

error[E0560]: struct `Transcript` has no field named `read_seq`
  --> src/align/stitch.rs:2158:9    read_seq: read_seq.to_vec(),
error[E0609]: no field `read_seq` on type `Transcript`
  --> src/align/stitch.rs:2698:28   transcript.read_seq = read_seq.to_vec();
  --> src/align/read_align.rs:892   t1.read_seq = mate1_seq.to_vec();
  ...

Not one read site, in src or in tests. The three places that looked like readers on a grep are something else: io/sam.rs:265 uses a local parameter of the same name, and chimeric/output.rs:148,157 read ChimericAlignment::read_seq, a different struct filled from its own parameter in chimeric/detect.rs.

What it costs

Measured on 200 000 real reads (ERR12389696, yeast) at 8 threads, with the global allocator wrapped in a counter:

	allocations	bytes
before	200 093 070	401 267 330 431
after	184 966 576	397 400 813 793

15.1 million allocations removed, 7.6% of the total, and 3.87 GB of copying. About 50 transcripts are finalised per read, each copying the whole read.

Wall clock: no change, and that is the interesting part

Six interleaved rounds, 2M reads, --runThreadN 16, 87-92% CPU idle:

	median
before	20.90 s
after	20.65 s

Direction mixed, inside the run-to-run spread. One round had a clear outlier which I have not excluded from the medians.

So removing 7.6% of all allocations moves nothing measurable. mimalloc is fast, and it is already tuned in src/main.rs. That calibrates #168 downward: I estimated an ~11% ceiling for allocation work from the profile's mimalloc share, and this says that estimate is too generous. I would not spend a new dependency on the rest of it without a workload that shows the cost, and I have updated #168 accordingly.

The reason to take this change is that the field is dead weight, which CONTRIBUTING.md rules out on its own terms. The allocations are a bonus, not the argument.

One thing needing your call

Transcript::read_seq is pub, so this is an API removal. Nothing outside the crate can be relying on its contents, since it is only ever written, but the field disappearing is still a breaking change for anyone naming it. Say the word if you would rather keep it as a deprecated always-empty field for a release.

Verification

Output-neutral: SAM byte-identical on 200 000 real reads against origin/main, whole file including header.

Gate: 560 lib + 26 integration tests, cargo clippy --all-targets -- -D warnings, cargo fmt --check, all green.