feat: add configurable embedding dimensions for self-hosted providers

[kaitranntt]

Description

Embedding dimensions are hardcoded to 1536 (OpenAI standard) in both the pgvector schema and provider wiring. Self-hosted providers like Ollama with models that output different native dimensions (e.g. qwen3-embedding:8b at 4096) cannot be used for embeddings even though:

1. The WithDimensions() plumbing already exists on OpenAIEmbeddingProvider
2. Ollama's OpenAI-compatible endpoint supports the dimensions parameter for Matryoshka truncation
3. The DB registry fallback path (createEmbeddingProvider → registry) creates providers without WithDimensions()

Root Cause

Commit 4db5259 added WithDimensions(1536) only for the Gemini case. The registry-based provider path (commit 1b99406) doesn't call WithDimensions() because all initial providers (OpenAI, OpenRouter) natively output 1536.

Affected Files

• internal/config/config.go — MemoryConfig struct missing EmbeddingDimensions field
• cmd/gateway_agents.go — createEmbeddingProvider() registry path doesn't apply dimensions

[mrgoonie]

Good catch on the hardcoded 1536 dimensions. The WithDimensions() plumbing exists but isn't wired through to the config path - that's a gap.

For self-hosted providers like Ollama where models output different native dimensions, this needs to be configurable. Your root cause analysis (the registry-based path not calling WithDimensions()) is correct.

Would you be interested in submitting a PR for this? The changes are well-scoped:

1. Add EmbeddingDimensions to MemoryConfig
2. Wire it through createEmbeddingProvider() in cmd/gateway_agents.go

If so, go for it - we'd be happy to review.