Version 3.5.1: re-runs work, upgrades re-run only what changed

3.5.1 replaces 3.5.0, which was tagged but never published. Everything below
is relative to 3.4.0.

Adding a genome to a finished run works. In every earlier v3 release a
re-run into an existing output directory failed in three tools, each tripping
over what its previous run left behind: Panaroo refused its own resume
manifest, CheckM2 refused a non-empty directory, and GTDB-Tk reused its
previous ANI screen, so the new genome went to pplacer, which asked for 132 GB
and died. Panaroo's leftovers and GTDB-Tk's directory are now cleared first,
and CheckM2 runs with --force.

Upgrading from 3.4.0 re-runs only what changed: snp-dists and FastTree
(the results change, below), checkm2 (--force), GTDB-Tk (its new first step),
and CarveMe with biosynthesis (a fixed solver). Nothing else, and no database
is fetched again.

CarveMe has a solver again. A fresh 3.4.0 install resolves scip 10.1, which
its pyscipopt cannot load, so every carveme job failed with "No solver
available." scip is pinned >=10.0.3,<10.1.

Panaroo no longer aborts on Bakta's selenoprotein genes (issue #156), and
--set now reaches every tool; for seqkit, checkm2, gtdbtk, mlst, panaroo,
snp-dists and FastTree it was accepted and silently ignored.

Results change. snp-dists and FastTree read Panaroo's filtered core
alignment, the one Panaroo recommends for phylogenies: on the eight-genome
Streptococcus showcase, 11-21% fewer SNPs per pair and one of five tree
splits moved; on the four E. faecium test genomes, 68-81% fewer, and the
closest pair changed. Counts and trees from 3.4.0 are not comparable.

Stopping a run behaves. Ctrl-C waits for Snakemake to cancel its jobs and
release the lock, instead of leaving a lock that made the next run fail and
then report "some tools failed". A locked directory is refused up front, with
the --unlock command. Under --tui, both quit options act on the key and
nothing further starts; a failed run exits 1, as it does without --tui; and a
Snakemake older than 9.14, which the interface cannot follow, is refused.

Under a cluster profile, a failed job is retried once unless the profile sets
its own retries.

Refused up front, instead of failing later: an --output or --databases
holding anything but letters, digits and _ . / -; a directory given as an
assembly; --report-only over a directory with no results; and running tools on
macOS or Windows, where none can be installed (--dry-run and --report-only
still work there).