epic: document and design how to handle chunked responses

[julianstephen]

Problem

As responses from mcp server tools can be chunked or streamed we need to design how the plugin adapter will invoke guardrails. E.g., options can be:

1. Wait for full response and then invoke guardrails (increased user wait time, and in case of very long running tools can break functionality)
2. Call guardrails with all partial responses (depending on the guardrail, this may be ineffective in specific situations)

Goals and non-goals

• Goals: Document usability and technical tradeoffs, and howtos.

Success criteria

• tests that simulate chunked behavior for future features
• docs that describe expected behaviors

Stories to be associated with this epic

No response

Additional Context

No response

[julianstephen]

Some relevant links:

• ext_proc: FULL DUPLEX STREAMED response body truncated when multiple filters defined envoyproxy/envoy#41654