Add async TEE resolution and public refresh-loop control

[adambalogh]

Summary

Backend relays (chat-api's OHTTP relay is the motivating case — see OpenGradient/chat-api#179) resolve a TEE on every request from an async server, but the SDK only offered a sync path with two problems:

1. resolve(tee_id) scans the on-chain registry with a blocking web3 call whenever the requested TEE isn't the active one — stalling the caller's event loop — and it rescans on every call.
2. The background TEE failover loop only starts lazily from chat/completion, so a server that only relays OHTTP ciphertext never gets failover: the active TEE resolved at startup stays pinned until restart, even after that TEE is upgraded and shut down.

chat-api currently papers over both with its own wrapper (thread dispatch + TTL memo + a private _tee.ensure_refresh_loop() poke). This PR moves that into the SDK where it belongs.

Changes

• aresolve(tee_id) on TEEConnectionInterface / LLM.aresolve_tee_connection(tee_id): event-loop-safe resolve for per-request use.
  • Registry scans run in a worker thread (asyncio.to_thread) behind a lock, so concurrent cold lookups collapse into a single scan.
  • Each pinned id's outcome — found or not-active — is cached for 30s, so steady traffic costs at most one chain RPC per TTL window per id, and hammering with a bogus id can't turn into one RPC per request.
  • Expired entries are swept once the cache outgrows any plausible number of real TEEs, so relays forwarding untrusted browser-sent ids can't grow it without bound.
  • Also starts the background refresh loop, so long-running relays fail over automatically.
  • StaticTEEConnection.aresolve returns the static connection (no I/O), matching resolve semantics.
• LLM.ensure_tee_refresh_loop(): public starter for the failover loop, for servers that never call chat/completion.
• The refresh loop's own registry scan and reconnect() now run their blocking web3 calls in a worker thread too (previously they stalled the host's event loop every 5 minutes), and a failed reconnect logs a warning with an accurate message (the old debug message claimed a client-close failure it never checked).
• resolve_tee_connection docstring now warns async servers off the sync path.

No behavior change for existing sync callers.

Testing

• Full suite: 175 passed, 1 skipped (pre-existing skip)
• New TestAresolve coverage: no-pin/matching-pin fast path does no scan; pinned id scans once then caches; unknown id raises ValueError with the miss cached (single scan across repeated calls); 10 concurrent cold lookups collapse to one scan; refresh loop auto-starts; static connection resolves without I/O
• New TestTeeConnectionApi coverage for the LLM-level delegations
• ruff format and mypy src clean

🤖 Generated with Claude Code

https://claude.ai/code/session_01Wo1obT6mfb81hUDUekHDiK

---

Generated by Claude Code