feat: add retry logic with exponential backoff for embedding providers

[RobertLD]

Network calls to Ollama and OpenAI APIs have no retry logic — transient failures (timeouts, 429s, 503s) immediately fail the entire operation.

Suggested implementation:

• Add configurable retry count (default 3) and exponential backoff
• Retry on network errors, 429 (rate limit), and 5xx responses
• Log each retry attempt at warn level
• Apply to both embed() and embedBatch() methods