mirpy 3.18.0

BREAKING — mir signature emits the geometry half only

One tool per half. mir signature emits rsig, vdjtools signature emits vsig, and the caller
joins them on sample_id. Previously mir signature emitted both, which meant it spent ~94% of
its runtime recomputing a half the other tool already owns.

	vdjtools signature	mir signature
half	vsig — statistics	rsig — geometry
standard tier	160 columns	528 columns
channels	div (Hill/Shannon), clon, depth, len, pgen, iso, shm, pair, aa, pchem, kmer, qc, mask	div (Rao), depth, band, contrast, phiv, phij, phic

Measured at tier="standard" on one 4,000-clonotype sample: 10 ms for rsig against 128 ms for
both. mirpy's half is 77% of the columns for 8% of the cost.

mir signature      cohort/*.tsv.gz -o rsig.parquet     # geometry
vdjtools signature cohort/*.tsv.gz -o vsig.parquet     # statistics

The join is exact, not approximate. The scale reference standardises per column, so a
column's value is identical whether or not the other half was computed beside it — pinned by a
test comparing rsig_cohort against the same columns of signature_cohort.

There is no combined mode and no --half flag, not even to opt back in. A second way to ask
for the other tool's output would be a duplicate, and it is the cost this change removes.

Added — rsig_cohort, and one shared pool

• mir.signature.rsig_cohort — a cohort of this tool's half. What the CLI calls.
• mir.signature.signature_cohort — unchanged in meaning: the joined vector, which is what
  the scale-reference fitting machinery measures over. Not used by the CLI.

Both, and vdjtools.signature.vsig_cohort, now run on one implementation —
vdjtools.signature.cohort.parallel_rows (new in vdjtools 3.15.0). Three callers, one pool, no
copies of the process handling. Requires vdjtools>=3.15.0.

Presets are intersected, not ignored. mir signature --preset classify keeps the 514 rsig:
columns of that preset's 615 and says on stderr how many it dropped and which command emits them;
vdjtools signature --preset classify keeps the other 101. That mirrors behaviour vdjtools
already had. A preset with no rsig: columns is an error here, naming the other command, rather
than an empty frame.

Added — the run says what it chose, and the docs say what it costs

mir signature prints the sample count, tier, worker count and the cores it believes it has
before starting, then the elapsed and per-sample cost. The core count is the one thing a cluster
or container silently gets wrong, so it goes where somebody can see it and stop early.

[mir] 1000 samples, tier=core (preset compact), 8 workers of 8 available cores
[mir] 1000 samples x 86 columns (compact, standardize=reference) in 116.5 s (117 ms/sample)

:doc:signature gained a sizing section with end-to-end measurements on 1,000 samples of 10,000
clonotypes, including a clean pip install on an 8-core / 32 GB Linux node (Xeon Silver 4210R):
117.8 s at --preset compact, 274 MB peak, 774% CPU. Two things that table exists to say. The
tier dominates, not the column count — classify (615 columns) and transfer (550) measure
identically, because selecting fewer columns does not compute less. And per-core speed matters
more than core count: the same sample is ~1.0 CPU-second on an M-series core and ~4.0 on that
Xeon.