feat(core): named/shared schemas + flatten response allOf

[Joehoel]

Emit each OpenAPI definition once as a named/shared schema instead of inlining it at every $ref, then flatten allOf in response schemas so create/get operations return the full provisioned resource.

Stacked on #341 (the Azure LRO PR) — the first three commits are that PR and drop out once it merges.

Why

The generator inlined every $ref at each use site. For services that concatenate many operations into one file (Azure: ~50 ops in storage.ts), inlining a deep resource schema per operation made tsgo blow up — so response schemas had to be left as { id, name, type } stubs, which in turn meant the LRO poller resolved the full resource only for the output schema to strip it back.

What changed

• Named/shared schemas. Each definition is emitted once as a const (in a shared _schemas.ts; inlined once into the service file for Azure's concatenated layout) and referenced via Schema.suspend(() => XSchema). Cyclic back-edges degrade to Schema.Unknown (same as the existing inline seenRefs guard) — which also breaks the type-level recursion, so no hand-written Schema.suspend annotations are needed. Each schema is type-checked once.
• Response allOf flattening. With inlining gone, generateOutputSchema now flattens the top-level allOf inheritance chain, so outputs carry the resource's own + inherited properties:

- StorageAccountsCreateOutput = { id, name, type, systemData }
+ StorageAccountsCreateOutput = { id, name, type, sku, kind,
+   properties: { provisioningState, … }, … }   // via shared StorageAccountPropertiesSchema

Regenerated Azure (27M, smaller than the prior 32M thanks to dedup) and Neon. Other OpenAPI packages keep their committed output until next regenerated.

Validation

• tsgo clean locally on Neon (the per-operation path) — proves the named-schema + suspend mechanism compiles.
• Azure: oxlint 0 errors, oxfmt --check clean, and the live storage smoke test now asserts provisioningState: "Succeeded" end-to-end.
• Azure's full tsgo is large; it runs in CI here (the deduped output is smaller than the baseline that already passed CI).