`buildExportsSourceMap` scanned every generated line for each export
(O(exports x lines)), which dominated source-map generation for
export-heavy CSS modules (e.g. Tailwind as a module). Resolve each
export's line in one pass by matching each line's leading JSON key
against the requested export names.

…shing

`_escapeIdentifier` ran two regex tests per character and concatenated
char-by-char. Classify each ASCII code once via a lookup table and flush
safe runs in bulk, so identifiers that need no escaping return unchanged
with no allocation. ~11x faster on escape-heavy CSS-module class names
(e.g. Tailwind variant selectors); output is byte-for-byte identical.

…x scan

`_unescapeIdentifier` appended char-by-char and `gobbleHex` sliced +
lowercased a substring per backslash just to test the next char. Scan
hex by index (no slice, case-folded inline) and flush safe runs in bulk;
use `indexOf` as the fast path and start offset. ~2.5x faster on
escape-heavy selectors (e.g. Tailwind `focus\:sr-only`); output is
byte-for-byte identical.

The comment callback eagerly computed line/column (two `LocConverter.get`
calls + a loc object) for every comment, but only magic-comment error
warnings ever read it. Defer it behind a getter on a `Comment` class,
matching the existing lazy `Node.loc`. ~16% faster comment parsing on
comment-heavy CSS; output unchanged.