llm 0.33 upgrades OpenAI Python library to 3.x and replaces httpx with httpx2

The upgrade to OpenAI Python library 3.x and the switch from httpx to httpx2 are breaking changes for any downstream code that pins exact versions or relies on specific HTTP client behavior. Teams using llm in CI pipelines or automated workflows will need to test their scripts against the new dependencies, particularly if they interact with custom endpoints or proxies. The HTTP client change may also affect connection pooling, timeouts, or error handling, which could surface in edge cases like rate limiting or network interruptions.

Per-call API key support for embeddings introduces a more flexible way to route requests to different providers without modifying the global configuration. This is useful for multi-tenant setups or experiments where different keys are needed for different embedding models. Existing plugins that read `self.key` will continue to work due to a compatibility fallback, but new plugins should use the `key=` parameter to avoid future deprecation. The change does not affect regular LLM model calls, which already supported per-call keys.

Template chaining via repeated `-t/--template` flags allows combining model configurations and prompts from separate templates. This reduces duplication in workflows where the same model settings are reused across multiple prompts. The reasoning summary option for OpenAI Responses API models provides a way to control the verbosity of outputs, which is particularly useful when testing or comparing models that mimic the OpenAI API. These features are additive and do not require immediate adoption but may simplify complex workflows over time.