Epic: LLM provider cleanup

[bug-ops]

Summary

LLM provider layer has a memory leak (Box::leak), code duplication (3x retry loops, SSE construction), unnecessary allocations (double clone in embed_fn), and inconsistent HTTP client creation.

Child Issues

• Change LlmProvider::name() from &'static str to &str #633 Change LlmProvider::name() from &'static str to &str
• Extract rate-limit retry helper in zeph-llm #634 Extract rate-limit retry helper
• Extract sse_to_chat_stream helper #635 Extract sse_to_chat_stream() helper
• Eliminate double AnyProvider::clone() in embed_fn() #636 Eliminate double AnyProvider::clone() in embed_fn()
• Share single reqwest::Client across provider instances #637 Share single reqwest::Client across provider instances
• Cache JsonSchema per TypeId in chat_typed #638 Cache JsonSchema per TypeId in chat_typed