Feat/shoaib annotation consistency

[Shoaib-Altaf]

What does this PR do?

Implements check_annotation_consistency — Week 4 deliverable, flags samples with
annotation confidence below a configurable threshold.

Which module(s) does it touch?

inference_audit/checks/annotation_consistency.py (new)
tests/test_annotation_consistency.py (new)

How did you test it?

Ran: pytest tests/test_annotation_consistency.py -v

============================= test session starts ==============================
collected 10 items

test_high_confidence_scores_high PASSED
test_low_confidence_scores_in_expected_range PASSED
test_no_confidence_column_provided_is_graceful_skip PASSED
test_missing_column_returns_none_score PASSED
test_empty_dataframe_returns_none_score PASSED
test_invalid_threshold_above_one_returns_none PASSED
test_invalid_threshold_negative_returns_none PASSED
test_null_confidence_values_reported_explicitly PASSED
test_all_null_confidence_returns_none PASSED
test_non_numeric_values_returns_none PASSED

============================== 10 passed in 0.47s ==============================

Also ran against both real fixtures: high_confidence.csv scored 100.0 (expected 90-100), low_confidence.csv scored 45.4 (expected 30-50) — both landed cleanly inside the Role Guide's expected ranges, no discrepancy to flag on this check.

Edge cases handled:

• conf_col=None (not provided) -> graceful skip, CheckResult(score=None, ...) — this is the expected normal case for datasets without confidence data, not an error
• Missing column -> CheckResult(score=None, ...)
• Empty dataframe -> CheckResult(score=None, ...)
• confidence_threshold outside [0, 1] -> rejected with a clear explanation, since a threshold outside that range is meaningless for confidence values
• Null values in the confidence column -> computed and reported explicitly in details and the warning (excluded from the rate calculation denominator, never silently dropped without a trace) — applying the same discipline the director's review
  required for check_label_distribution's null-label handling
• All confidence values null -> CheckResult(score=None, ...), nothing to assess
• Non-numeric values in the confidence column (e.g. strings) -> rejected with a clear explanation rather than silently miscompared against a numeric threshold — applying the same discipline the director's review required for check_missing_values' non-string handling

Checklist:

• Tests added or updated
• Docstrings on all new public functions
• No unused imports
• README updated if public API changed (N/A — no public API change)
• Lead Engineer has reviewed and approved before submitting to director