LLM client code refactor

[adambalogh]

Summary

Refactors the LLM client to use async/await and introduces the TEE Registry for dynamic endpoint discovery.

• Async LLM client: All LLM request methods (chat, completion) are now async, with streaming using AsyncGenerator instead of threaded TextGenerationStream
• TEE Registry integration: LLM endpoints and TLS certificates are now discovered on-chain via the TEERegistry contract instead of hardcoded server URLs. The pinned TLS cert from the registry replaces TOFU trust
• TEE metadata in responses: TextGenerationOutput and StreamChunk now include tee_id, tee_endpoint, and tee_payment_address fields for full provenance
• Removed og_llm_streaming_server_url: Unified behind a single optional og_llm_server_url override, with registry-based discovery as the default
• Removed custom exceptions.py: Cleaned up unused exception classes
• Simplified Makefile: Consolidated individual test targets into a single make test that runs all tests via pytest tests/
• New llm_test.py: Comprehensive unit tests for the async LLM client
• Updated examples and docs: All examples and generated docs reflect the async API and TEE registry changes

Test plan

• make test — all unit tests pass
• make check — formatting and type checks pass
• make integrationtest — integration tests pass against devnet
• Manual: make chat and make chat-stream work with live TEE endpoints