syntax: fix class negation across the surrogate gap

[CarloTUM]

Fixes #1397.

IntervalSet::negate computes the gap between two adjacent canonical ranges with increment()/decrement(). For ranges on either side of the surrogate codepoints, such as ..=\u{D7FF} and \u{E000}.., both steps jump over the surrogates, which yields lower = \u{E000} and upper = \u{D7FF}. Interval::create then swaps the inverted bounds and produces \u{D7FF}..=\u{E000}, so both codepoints end up matching their own negation:

Regex::new(r"^[^\x{D7FF}\x{E000}]$").unwrap().is_match("\u{D7FF}") // true

canonicalize never merges these two ranges, since is_contiguous compares the integer values and 0xD7FF + 1 != 0xE000. So the fix handles the empty gap in negate itself: when lower > upper there is nothing between the ranges and no range is pushed. Byte classes are unaffected, as u8 has no such gap.

Tests:

• unit tests in hir/mod.rs for the three shapes: the two codepoints alone, the full set whose negation must be empty, and a mixed set,
• regression tests in testdata/regression.toml for U+D7FF, U+E000 and the neighbouring U+D7FE.

I left the CHANGELOG untouched; happy to add an entry if you prefer that in the PR.

---

Disclosure: assisted by GitHub Copilot and Claude (including this description). I reviewed and tested the change myself.