[quentindemetz]

SchemaId is constructed for every sub-schema and reference, and each construction calls the legacy url.resolve / url.parse APIs, which are slow. Large OpenAPI documents create many SchemaIds, so this PR avoids most of that work in two ways:

1. Memoize url.resolve and url.parse results, since many SchemaIds share the same inputs.
2. Skip url.resolve for simple references. Most calls resolve a fragment (#/components/...) or an absolute path (/properties/...) against a base with no scheme or host. For those, the result is just the base without its fragment followed by the fragment, or the path itself. The shortcut only applies when both strings contain letters, digits, -, _, ., ~, / and # only, no . or .. segments, and do not start with //. Everything else (percent-escaping, dot-segment removal, URLs with a host) still goes through url.resolve.

On a codebase generating types for ~2,000 OpenAPI files:

• memoization cut url parsing from ~10.6 s to ~3.2 s of CPU time (−30% of total generation time);
• the shortcut then cut total type generation from 8.1 s to 4.7 s of CPU time (−42%). 154,479 of the 154,504 distinct url.resolve calls take it, and every one returns the same string as url.resolve.

Output is byte-identical. New tests cover both the shortcut and the cases that must fall back to url.resolve.