Scope extension-relation schema derivers to the ExtensionRegistry instead of a process-global

[nielspardon]

Follow-up cleanup from review of #204 (non-blocking; not a correctness bug).

substrait.api's user-defined extension relations register their schema deriver in a process-global dict rather than on the ExtensionRegistry instance:

• type_inference.register_extension_relation writes to the module-global _extension_relation_derivers (src/substrait/type_inference.py), keyed by type_url.
• ExtensionRegistry.register_extension_relation also writes self._extension_relations[...], but that per-instance dict is never read by inference — it's effectively dead state.

Consequences:

• Two ExtensionRegistry instances silently share one deriver map; registering a detail class on one changes schema derivation for plans built with another.
• Last-writer-wins if two detail classes share a type_url.
• There is no way to unregister or isolate — state leaks across otherwise-independent registries (and across tests).

Every other registry capability is scoped to the instance; extension-relation derivation should be too. Suggested direction: store derivers on the ExtensionRegistry and have _derive_extension_schema consult the registry already threaded through inference, then drop the global and the dead instance dict.

🤖 Filed with AI assistance as a follow-up to #204.