Fix byte-string Unicode semantics for regex word classes

[fglock]

Summary

PerlOnJava regressed byte/UTF-8-sensitive regular-expression behavior exercised by Test::Regexp.

In the Test::Regexp UTF-8 matrix, cases involving the byte form of F\x{F8}o and \w+ report a successful match when the upstream test expects no match. This causes follow-on failures for match completeness and produces 36 semantic assertion failures in t/140_utf8.t.

CPAN impact

Test::Regexp 2017040101 regressed in CPAN tester run 20260828-172308-14550 after its prior recorded pass on cc69cbb4c (2026-08-12).

The run fails 55 of 2,253 subtests:

• t/140_utf8.t: 46 failures, of which 36 are the byte/UTF-8 regex-semantic failures.
• t/120_keep_pattern.t: 9 named-capture pattern-rendering failures. Those are tracked separately from the semantic problem here because capture results themselves are correct.

Oracle and scope

System Perl 5.42 passes the complete unchanged distribution suite: 23 files and 2,251 tests (with its ordinary author-test skips).

Both PerlOnJava JVM and interpreter backends reproduce t/140_utf8.t failures. The failures occur through Test::Regexp::match while it varies UTF-8 flag state with utf8::upgrade and utf8::downgrade; they are not a source-filter or native dependency problem.

The byte-mode capture corruption tracked in #1183 may be related at the representation boundary, but this case is a direct \w matching/flag-semantics regression and needs an independent minimal reproducer before sharing an implementation fix.

Acceptance criteria

• Add a focused project-owned regression based on the smallest system-Perl-validated \w case that distinguishes upgraded and downgraded Latin-1 strings in the same way as Test::Regexp.
• Match system Perl's result and match-completeness behavior on JVM and interpreter backends.
• Preserve existing Unicode behavior for UTF-8-flagged strings.
• Test::Regexp 2017040101 t/140_utf8.t passes on both backends.