FIX: Include matcher configuration in decoding scorer identity

[Ayoub (Ayoubhm07)]

Description

Closes #3074 .

DecodingScorer instances with different matcher settings could produce opposite verdicts but share both hash and eval_hash, allowing cached evaluation metrics to be reused across different behavior. This is the same defect #2977 described for SubStringScorer, in the sibling class that #2979 did not touch: both hold a TextMatching instance at the same point of the same method, but only SubStringScorer._build_identifier() passes its get_identifier_params() through.

The fix copies the two lines #2979 added to the sibling, so the built-in matchers' case sensitivity, whitespace handling, threshold, and n-gram size now reach the identity. Custom matchers that only implement is_match() remain supported through the same callable() guard. Built-in matcher hashes change intentionally, as they did for SubStringScorer; no shipped scorer metric registry references DecodingScorer, so no published eval_hash moves.

Scorer categories reach the emitted score_category but are absent from both scorers' identifiers. #2979 left them out for SubStringScorer too, so I kept this change symmetric with the sibling rather than widening it; happy to cover categories for both in a follow-up if you want them included.

Tests and Documentation

• Added the sibling's two identity regressions, adapted to this scorer: three parameterised cases covering case sensitivity, threshold and n-gram size, each asserting that opposite verdicts no longer share an identity, plus equivalent-configuration stability for both built-in matchers.
• Before the fix: 3 failed, 11 passed in tests/unit/score/test_decoding_scorer.py. After: 14 passed.
• The full tests/unit/score/ suite reports the same 47 pre-existing failures before and after, with passes going from 3422 to 3427, which is exactly the five added cases. Those failures are confined to test_azure_content_filter.py and test_local_refusal_classifier_scorer.py and are unrelated to this change.
• ruff check and ruff format --check at the pinned v0.16.10, git diff --check, check_async_suffix.py and check_no_rest_roles.py all pass. ty check reports the same single pre-existing diagnostic on line 50 before and after, which substring_scorer.py also reports. I ran pytest directly rather than make unit-test, and did not run the full repository suite.