Support multiple ITSx profile sets with per-read best-match selection (-t all equivalent)

[ayobi]

Requested by @vmikk in #5. Splitting it out since the SSU/LSU half of that issue shipped in v0.3.0 and this part needs its own design discussion.

What's wanted

ITSx's -t all searches every taxonomic profile set and reports a preliminary origin per sequence. ITSxRust currently takes a single --hmm file, so a mixed eukaryotic community amplified with universal primers gets delimited against one profile set regardless of what each read actually is.

Design notes

A single concatenated search looks preferable to N separate nhmmer passes. ITSx profile names already carry the taxonomic group:

1_SSU_Fungi_Ge_1_SSU_end_five_long
2_5.8_Fungi_Ge_58S_start_onee_long
3_End_Fungi_Ge_58S_end_oneg_long
4_LSU_Fungi_Ge_4_LSU_start_five_long

So one pass over a merged HMM file would let select::classify bin hits by both anchor type and group, at roughly the cost of one search rather than one per set.

Open questions:

• Selection criterion. Summed anchor bitscore across the four anchors is the obvious choice, but profile sets differ in model count and model length, so a set with more or longer models may win systematically. This needs checking against sequences of known origin before it can be trusted.
• Ties and mixed evidence. What to do when two sets score comparably, or when the SSU anchor favours one group and the LSU another. The latter is a chimera signal worth surfacing rather than silently resolving.
• Reporting. The winning group should appear per read in the anchors TSV and be summarised in the QC JSON, consistent with the existing diagnostics.
• Memory. Top-K is currently tracked per anchor type per strand. Adding a group dimension multiplies the accumulator, which matters now that peak memory scales at roughly 10 kB per read.
• Backwards compatibility. A single --hmm file should keep working unchanged; multi-set behaviour should be opt-in.