[DylanCosto]

The quoted_semicolon_patterns regexes split on a separator only if the rest of the string has balanced quotes. That lookahead rescans the remaining text at every separator, so parsing is quadratic in the attribute length. It gets slow for GTF records with many attributes or quoted values that contain semicolons.

This replaces the regex with _split_on_unquoted_separator, which splits naively and then rejoins pieces until the number of quotes to the right is even. That's the same condition the lookahead checks, including for unbalanced quotes, but each piece is scanned once.

Rough timings on synthetic files:

• iterating 2,000 GTF records with 40 attributes: 1.36 s → 0.15 s
• building an in-memory db from 200 GTF records with 200 attributes: 4.6 s → 0.11 s
• GFF3 records with only a few attributes: about the same

[daler]

This is great, thanks! Will target v0.15rc with this.

xref #212 #215