Fix UTF-8 false-positive on short probes with tolerated errors (regression from #2882 + #2878)

[tballison]

Summary

Fixes a CI regression introduced by the interaction of #2878 and #2882:

• TIKA-4745 -- efficiency improvements #2878 added a utf8Tolerated path: when a probe is NOT_UTF8 but has ≤ UTF8_MAX_TOLERATED_ERRORS error events, the UTF-8 NB result is kept (not disqualified) and a STRUCTURAL UTF-8 candidate is also emitted.
• make mojibuster default #2882 changed the return strategy: when the top result is STRUCTURAL, return only that one result and skip JunkFilter arbitration.
• Combined effect: A short probe like "the name is \u00e1ndre" in ISO-8859-1 (17 bytes, 1 UTF-8 error event) was being promoted to STRUCTURAL UTF-8, which then short-circuited JunkFilter, preventing it from comparing UTF-8 vs windows-1252. Result: UTF-8 returned instead of windows-1252.

Fix: Only emit the STRUCTURAL UTF-8 candidate for LIKELY_UTF8 (definitively clean grammar). For utf8Tolerated (NOT_UTF8 but within tolerance), NB's UTF-8 result is already kept as a STATISTICAL candidate — JunkFilter has enough signal to arbitrate correctly between UTF-8 and windows-1252.

Failing tests fixed: TXTParserTest.testUseIncomingCharsetAsHint and TXTParserTest.testUsingCharsetInContentTypeHeader.