[aknight-ld]

Add in a max retries passthrough argument to let handler factories natively retry.

---

Note

Overview
Adds an optional max_retries knob on the messages handler factories (Claude, OpenAI, LangChain, LiteLLM) and their convenience wrappers, so callers can raise provider-level retries for bursty workloads (e.g. concurrent eval workers) without touching model.parameters.

Shared validate_max_retries in launchdarkly_ai_server enforces None or a non-negative int at handler construction (including rejecting bool). Omitted values leave each SDK’s default unchanged; model.parameters cannot override the handler setting.

Wiring differs by integration: native Anthropic/OpenAI clients get max_retries only when set; LangChain applies it to auto-built chat models and errors if combined with a custom llm; LiteLLM picks max_retries vs num_retries per route (OpenAI/Azure vs everything else) so retries stay on one layer.

^{Reviewed by Cursor Bugbot for commit 4bed864. Bugbot is set up for automated code reviews on this repo. Configure here.}