feat(grma): generalize to k loci + compact tree storage

[Ofekirsh]

Summary

This PR delivers four changes:

1. Correct GVH/HVG and mismatch computation.
2. Fix a missing-donor bug in the candidate list.
3. Generalize to k loci (no longer limited to 5).
4. Support additional allele formats (beyond xy:wz), while keeping the genotype representation compact.

---

What changed

(1) GVH/HVG and mismatches

• For each locus, patient/donor alleles are handled as sets (no duplicates).

• We now compute sizes (cardinalities):

  • GVH := | Patient \ Donor |
  • HVG := | Donor \ Patient |

• The number of mismatches is max(GVH, HVG).

File: grma/match/donors_matching.py.

---

(2) Missing donor in candidates

• In cpdef tuple neighbors_2nd(self, UINT node), a duplicate -1 placeholder was present; removing the extra one prevents dropping a donor from the results.

File: grma/match/lol_graph.pyx.

---

(3) k-loci generalization

• Removed hard-coded assumptions (e.g., the magic number 10) and replaced them with configurable values.
• Verified for 4/5 loci; full 6-loci coverage wasn’t targeted here due to edge cases (e.g., DRB3/4/5, DRBX last locus, etc.).

---

(4) Additional allele formats + compact storage

• Broadened parsing/handling beyond xy:wz (varying digit widths/fields).
• To avoid donor-tree bloat, we use UIDs per allele via a bidirectional map (bidict).
• Classes/subclasses use hash keys for compact indices; any hash collisions are harmless since exact mismatch checks operate on the UID layer.

Affected: many files (tree maintenance, matching, LOL/SLUG building, utilities).
New dependency: bidict.

---