Parse NIC Classes 2, 3 and 4 from OBR EFO Table 3.4

[MaxGhenis]

Summary

policyengine_uk_data/targets/sources/obr.py::_parse_nics only covered Class 1 Employee and Class 1 Employer NICs, omitting:

• Class 2 (flat-rate self-employed)
• Class 3 (voluntary)
• Class 4 (self-employed profit-based)

That's ~£2-5B/yr of NIC receipts with no calibration target, which biased the model into pushing self-employment income downward to compensate and distorted reforms touching SE income.

This PR:

• Adds rows for Classes 2, 3 and 4 in _parse_nics, emitting Targets with the matching policyengine_uk variable names (ni_class_2, ni_class_3, ni_class_4).
• Switches the per-row lookup to a list of candidate labels so minor OBR wording variation (e.g. "Class 2 NICs" vs "Class 2 Self-Employed NICs") no longer silently skips a row.
• Partial matches still produce useful targets — missing rows log a warning instead of aborting parsing.

Adds test_obr_nics.py, which drives the parser against an in-memory openpyxl workbook that mirrors OBR's Table 3.4 layout (canonical labels + the common wording variants).

Test plan

• uv run pytest policyengine_uk_data/tests/test_obr_nics.py -q passes.

Finding U8 from the bug hunt.

[MaxGhenis]

Self-review: APPROVE. See /tmp/bug-hunt/reviews/uk-data-reviews.md for detail. (GitHub blocks self-approve, so leaving as comment.)