meta: fix non-leftmost match in reverse-suffix/inner optimizations

[stefanobaghino]

Closes #1345.

ReverseSuffix::try_search_half_start and ReverseInner::try_search_full returned on the first successful reverse search, which assumed that the leftmost occurrence of the suffix (resp. inner) literal corresponds to the end of the leftmost-first match. That assumption breaks when the regex prefix is non-monotonic — e.g. an optional group whose body can absorb the literal between consecutive occurrences — so a strictly later literal occurrence may have a strictly earlier overall match start. The minimal repro filed in #1345 is [^()]*(?:\([^()]*\))?[^()]*: on $(:):: META was returning 2..3, while the leftmost-first match is 0..5 (the optional group absorbs (:) and pins to the second :).

Both loops now track the best (smallest-start) candidate and bound each subsequent reverse search by best.offset(). The existing limited reverse DFA in regex-automata/src/meta/limited.rs returns RetryError::Quadratic the moment scanning past the bound is required, which propagates through ? and triggers the existing Core fallback in Strategy::search. The loops also early-exit when the candidate already starts at input.start(), so the common (monotonic-prefix) cases stay one-shot.

The bug was reduced from real-world .sublime-syntax grammars driven through fancy-regex's seek-prefilter approximation; see fancy-regex/fancy-regex#249 for the downstream context.

testdata/regression.toml gets the minimal repro as non-monotonic-reverse-suffix; it failed against meta and passed against pikevm/hybrid before this change.