docs(self-hosting): clarify embedding provider env vars, local models, and v0.0.6 upgrade lock (#1336)

[Adityakk9031]

Resolves #1336

Summary

In self-hosted v0.0.6 and v0.0.7-rc.2, the compiled release binary assets were published without pluggable embedding configuration hooks (environment variables like SUPERMEMORY_EMBEDDING_PROVIDER, SUPERMEMORY_EMBEDDING_MODEL, and SUPERMEMORY_EMBEDDING_DIMENSIONS were not linked in the binary and defaulted to local English Xenova/bge-base-en-v1.5).

While full pluggable embedding support is linked in v0.0.7 and v0.0.8+, several documentation gaps and upgrade traps caused confusion:

1. Existing Data Directory Lock: When migrating from v0.0.6, pre-existing rows without an embedding-plan.json lock file cause the server to automatically lock to legacy local · Xenova/bge-base-en-v1.5 · 768d to prevent mixed-space vector corruption. Any newly configured embedding env vars are either blocked or fail boot unless the data directory is wiped or pointed to a fresh path.
2. Provider naming: SUPERMEMORY_EMBEDDING_PROVIDER only accepts local, openai, openai-compatible, or gemini. Setting SUPERMEMORY_EMBEDDING_PROVIDER=ollama causes an unknown provider error; Ollama requires openai-compatible (or openai) with SUPERMEMORY_EMBEDDING_BASE_URL.
3. Local Models & Native Dimensions: Clarified that local ONNX models do not support dimension reduction, and provided an explicit lookup table of all built-in local models (e.g., Xenova/bge-m3 at 1024d, Xenova/multilingual-e5-base at 768d, etc.) with required dimensions.
4. Added SUPERMEMORY_EMBEDDING_API_KEY to the configuration documentation.