Fix HTML literal recovery at EOF

[fglock]

Summary

• preserve HTML literal-element state across parse() chunks without emitting duplicate start events
• mirror HTML::Parser 3.85 EOF recovery for unterminated title content so following markup remains tokenizable
• add a standard-Perl-validated regression covering the exact token stream and networkless form count

Verification

• system Perl oracle: html_parser_literal_eof.t passes 2/2
• exact-master baseline: new fixture fails 2/2 on JVM and interpreter
• candidate: new fixture passes 2/2 on JVM and interpreter
• chunk-boundary reducer passes on system Perl, JVM, and interpreter
• networkless WWW::Mechanize reducer reports both substitutions successful and form counts 1,1 on both backends
• unmodified WWW::Mechanize 2.22 t/local/overload.t passes 9/9 on JVM and interpreter, including its loopback HTTP path
• full make: BUILD SUCCESSFUL in 5m 5s, 17/17 tasks executed, warning-free

Scope

This fixes PerlOnJava's HTMLParser.java; no WWW::Mechanize, HTML::Form, HTML::TokeParser, regex/Joni, or upstream test source is changed.