Add Letta memory provider

[ThePlenkov]

Summary

Adds a letta memory provider to the benchmark, backed by Letta's archive API for semantic retrieval and optional agent-based answers.

• Registers LettaMemoryProvider under --memory letta and regenerates catalog.json so the UI lists it.
• Ingests benchmark Documents as archival passages (one archive per isolation unit), batching writes in groups of 50 and embedding document timestamps in text to avoid date-format rejections.
• retrieve() searches the archive via passages.search, honors the caller's k (default 20), and filters out passages whose stored timestamp is after the optional query_timestamp to prevent future-context leakage.
• direct_answer() lazily creates a throwaway Letta agent per unit with the archive attached; prepends the question date to the prompt and serializes messages with a per-agent lock. cleanup() deletes temporary agents while keeping archives.
• Environment-based config: LETTA_API_KEY or LETTA_BASE_URL is validated at startup; optional LETTA_EMBEDDING_MODEL, LETTA_MODEL, LETTA_MAX_STEPS.

Adds letta-client>=1.12.1 to dependencies and documents the new options in README.md.

Verified with mocked Letta client only; a live run requires Letta Cloud or a self-hosted server.

[vercel]

@devin-ai-integration[bot] is attempting to deploy a commit to the Vectorize Team on Vercel.

A member of the Team first needs to authorize it.