refactor!: replace ModelRequestError with EmbeddingRequestError

[TheGreatAxios]

Summary

• Stacked on build(deps): move @intx/* to peers and require arktype ^2.2.3 #9.
• Every request failure (transport, HTTP, non-JSON 200, malformed/short/duplicate reply, bad base64, empty probe) throws EmbeddingRequestError extends Error with reason: InferenceError and url, classified by @intx/inference and retried via createDefaultRetryPolicy.
• ModelRequestError and the runJSONRequest/extractRetryAfterMs exports are removed; the transport is private.
• RequestDependencies and RetryAfterExtractor are exported as the types of EmbedOptions fields.
• options is optional; deps defaults to global fetch + createDefaultScheduler(), resolved once per embedTexts call.
• Config is asserted with EmbedConfigSchema (invalid config throws arktype's TraversalError); request objects are explicit literals.
• README Errors section, ARCHITECTURE.md and IMPLEMENTATION.md describe the new surface.

Verification

• typecheck, build and bun run test (20 pass); CI passes.

Closes CL-9052
Closes CL-9059
Closes CL-9050
Closes CL-9048
Closes CL-9054
Closes CL-9056

[TheGreatAxios]

Review

Independent review of the stack found: bad base64 replies escaped as a raw DOMException or silently truncated vectors; non-integer indices were accepted; duplicate-index and non-JSON-200 coverage was lost; the live suite failed instead of skipping when the model was absent. All fixed in this stack (strict base64 decode, number.integer index, restored harness tests, model-aware skip).

Remaining notes: peer range >=0.4.0 follows sibling convention. Each commit typechecks, tests, and builds standalone.