feat(ai-gateway): Document OpenAI format compatibility gaps across all providers [2.x]

The audit in `Kong_AI_Gateway___OpenAI__v1_chat_completions_Mapping_Status.pdf` shows each provider driver handles OpenAI `/v1/chat/completions` differently. Some fields work universally (temperature, messages), others silently drop (top_k for Anthropic/Bedrock, top_p for Anthropic, stop → stop_sequences conversion for Anthropic), and some cause upstream errors (vision inputs across Anthropic/Bedrock/Gemini). Cache metrics aren't collected for custom-transform providers. Developers using `llm_format: openai` assume field parity. They don't have it.

Add a cross-provider compatibility matrix to the AI Gateway docs showing which OpenAI fields are supported, dropped, incorrectly mapped, or cause errors per provider (OpenAI, Azure, Anthropic, Gemini, Bedrock, Cohere, etc.). Include workarounds where applicable. Direct language. Link to the audit for reference.

[Kong AI Gateway — OpenAI _v1_chat_completions Mapping Status.pdf](https://github.com/user-attachments/files/29276354/Kong.AI.Gateway.OpenAI._v1_chat_completions.Mapping.Status.pdf)

Start with the linked Kong AI Gateway — OpenAI /v1/chat/completions Mapping Status PDF and review each provider’s field behavior. Build a cross-provider compatibility matrix covering supported, dropped, incorrectly mapped, and error-causing fields, with applicable workarounds and a reference link to the audit. Done means the matrix covers the named providers and clearly documents the compatibility gaps.

Written by the indexing model from the issue text.