feat: add prompt caching schema support for system/developer messages

[dcox761]

*Issue #229 *

Description of changes:

Add CacheControl model and cache_control field on TextContent to support OpenAI-compatible prompt caching hints (e.g., cache_control: {type: 'ephemeral'}).

Changes:

• schema.py: Add CacheControl model; add cache_control field to TextContent; change SystemMessage.content and DeveloperMessage.content from str to str | list[TextContent] to accept structured content with cache hints
• bedrock.py: Update _parse_system_prompts to handle list-format content with cache_control, emitting Bedrock cachePoint blocks; update _parse_content_parts to emit cachePoint for user message content parts

This allows clients to mark specific content blocks for caching using the standard OpenAI cache_control field, which is translated to Bedrock's cachePoint tagged union format.

Ref: https://docs.aws.amazon.com/bedrock/latest/userguide/prompt-caching.html

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[zxkane]

Thanks for the PR! The SystemMessage.content: str | list[TextContent] change is a welcome fix — it aligns our schema with the official OpenAI API spec, which does support string | array of ChatCompletionContentPartText for system/developer messages.

I had a question about the cache_control part of this PR though. After checking the OpenAI API reference, it appears that cache_control is not part of the official ChatCompletionContentPartText schema — OpenAI's prompt caching is automatic and transparent (prefix-based, 1024+ tokens), with no client-side hints needed.

The cache_control: {type: "ephemeral"} pattern originates from the Anthropic API. In frameworks like LangChain, it's only used with Anthropic-specific integrations (ChatAnthropic, ChatOpenRouter for Anthropic models), not with ChatOpenAI. Since this gateway is designed to receive OpenAI SDK-formatted requests, standard OpenAI SDK clients wouldn't send this field.

Could you share more about the use case you have in mind? For example:

• Are there specific clients or frameworks that send cache_control in OpenAI-compatible format to this gateway?
• Is this intended for users who craft raw HTTP requests with mixed OpenAI/Anthropic conventions?

Understanding the real-world scenario would help us evaluate the best approach. If there is a valid use case, we'd want to make sure the cachePoint injection respects the existing ENABLE_PROMPT_CACHING flag and _supports_prompt_caching() checks to avoid bypassing the current caching governance.

Again, the content: str | list[TextContent] fix for system/developer messages is great regardless — happy to see that part land!