refactor(embedding): simplify to OpenAI + Ollama with O(1) character-based truncation

[jgpruitt]

Summary

• Replace CPU-intensive tiktoken tokenization with O(1) character-based truncation
• Simplify embedding providers to just OpenAI and Ollama
• Add retry logic for OpenAI context length errors

Changes

Truncation overhaul

• Removed js-tiktoken and mistral-tokenizer-js dependencies
• New approach: Character-based truncation using 3.8 chars/token ratio (~10% buffer)
• text.length and text.slice() are O(1) — no event loop blocking
• OpenAI retries with progressively tighter ratios (3.8 → 3.0 → 2.5) if estimate is insufficient

Provider simplification

• Removed Cohere, Mistral, and Google providers (and their SDK dependencies)
• Kept OpenAI (hosted platform) and Ollama (self-hosted option)
• Removes ~70 lines of provider-specific code

Options cleanup

• Wired up maxParallelCalls to actually be passed to embedMany
• Removed unused batchSize option (SDK handles chunking automatically)

Why

1. Performance: tiktoken tokenization is CPU-bound and blocks the event loop
2. Simplicity: We only need OpenAI (hosted) and Ollama (self-hosted)
3. Cost: OpenAI is cheaper than self-hosting until ~80K memories/day

Breaking Changes

• EmbeddingProvider type now only accepts "openai" | "ollama"
• Removed batchSize from EmbeddingOptions

Test Results

• 201 tests pass
• TypeScript compiles cleanly
• Biome lint passes