Fix backend CI: upgrade datasets, use canonical squad id (closes #18)

[elkaix]

The Backend tests CI check has been red since before #17 (merged via admin override) — the pre-existing #18 dependency incompatibility. This fixes it at the root.

Root cause

A three-way version conflict:

• requirements.txt capped datasets>=2.16,<3. The datasets 2.x load path resolves a bare squad_v2 dataset id via huggingface_hub.
• huggingface_hub floated to 1.x (transitively), which rejects bare ids — it requires namespace/name. So load_dataset("squad_v2") raises HfUriError on any machine without a warm cache (i.e. CI).
• Pinning hub <1.0 is not an option: transformers 5.x requires huggingface_hub>=1.5.0.

Fix

• requirements.txt: datasets>=2.16,<3 → datasets>=3 — datasets 5.x supports huggingface_hub 1.x, so the whole stack (datasets / hub / transformers / sentence-transformers) is coherent (pip check clean).
• src/eval/datasets/squad_v2.py: load the canonical rajpurkar/squad_v2 id, which hub 1.x accepts. Same underlying data and row shape — the checked-in frozen artifact and the sampling tests are unaffected.

Verification

• Full suite 304 passing on datasets 5.0.0 + huggingface_hub 1.23.0.
• The 5 previously-failing squad tests (test_eval_datasets_squad.py) pass from an empty HF cache — the exact cold-start condition that fails in CI.

Closes #18. Unblocks the required Backend-tests check so future PRs merge without admin override.