Fix/association query and drug scoring

[dalloliogm]

Three bugs in open_targets_client.py caused clinical scores to be systematically wrong — most visibly, well-validated targets like IL17A and ADRA2A scored 0 clinical evidence even for diseases where approved
drugs exist. The root cause was a combination of an invalid GraphQL argument, a broken API subquery, and a scoring function that ignored which disease a drug was actually approved for.

• get_association used a non-existent GraphQL filter (associatedTargets(Bs: $ensemblIds)) — the Bs argument is not part of the Open Targets API schema. Every query returned zero rows silently, triggering a
  fallback that fetched 500 targets and filtered in Python on every single call. Fixed by switching query direction to target → associatedDiseases(efoIds: [...]), which uses the correct documented filter and
  eliminates the fallback entirely.
• diseases subquery caused silent 400 errors — drugAndClinicalCandidates returns ClinicalDiseaseListItem objects, so the correct subquery is diseases { disease { id name } }, not diseases { id name }. The
  old query was rejected by the API, masking all drug data without surfacing an error to the user.
• stage_map used stale string values — the Open Targets API now returns APPROVAL, PHASE_3, PHASE_2_3, etc. instead of the legacy "Approved", "Phase III" format. No stage values were matching, so all drugs
  contributed a phase of 0. Both formats are now supported.
• score_clinical ignored indication when scoring drugs — max_phase was computed across all drugs for a target globally, regardless of whether they were approved for the queried disease. Drugs are now split
  into indication-matched (full phase value) and off-indication (capped at phase 2). This is why IL17A scored 0 for psoriasis despite secukinumab being approved for it.
• Added tests/test_scoring.py with 8 unit tests covering the fixed logic — no network calls, using hand-crafted fixture dicts matching the new API response shapes.