Add local model connector options for local server

[seagpt]

Summary

Adds first-class local model connector options for the local Supermemory launcher and Supermemory.local() startup path.

This keeps the local server's existing OpenAI-compatible model-provider interface, but makes common local runtimes explicit:

• CLI shortcuts for Ollama, LM Studio, llama.cpp, and generic OpenAI-compatible endpoints
• Supermemory.local({ localModel: ... }) TypeScript options for programmatic startup
• loopback-friendly dummy OPENAI_API_KEY fallback for local endpoints that ignore bearer tokens
• README examples and a client test covering the new local-model option shape

Why

The local server already supports OpenAI-compatible model routing through env vars such as OPENAI_BASE_URL, OPENAI_MODEL, OPENAI_FAST_MODEL, and OPENAI_TEXT_MODEL. In practice, users running fully local memory stacks need a documented, typed way to point Supermemory at Ollama / LM Studio / llama.cpp / vLLM without falling back to hosted LLMs or manual env glue.

Validation

Ran locally:

• node --check bin/cli
• env -u SUPERMEMORY_API_KEY corepack yarn test --runInBand tests/index.test.ts → 52 passed
• env -u SUPERMEMORY_API_KEY ./node_modules/typescript/bin/tsc
• env -u SUPERMEMORY_API_KEY corepack yarn build → success
• git diff --check

I also smoke-tested the patched CLI against a local llama.cpp/OpenAI-compatible endpoint via --llamacpp-model + --llamacpp-url; the local server started, accepted a document, used local embeddings, issued a /v1/responses request to the configured local endpoint, and search found the inserted test document.

Notes / known limitation

This PR only adds the connector/configuration surface. During local smoke testing, the self-hosted memory-agent path still failed after receiving a local Responses API function-call response, even though document add/list/search and local embeddings worked. That appears to be a deeper local-server/model-agent compatibility issue rather than SDK routing, and likely belongs in the server implementation rather than this SDK launcher patch.