feat: add Responses compaction support

[franciscojavierarceo]

Summary

• add OpenAI-compatible POST /v1/responses/compact, reusable compaction items, previous-response rehydration, persistence, and automatic threshold-based context compaction
• preserve canonical user messages, replace prior compaction checkpoints, translate local plaintext encrypted_content into assistant context for upstream inference, and accumulate compaction usage across gateway tool loops
• add reproducible OpenAI cassette recordings and offline replay coverage for standalone, repeated, round-trip, and automatic compaction paths using the existing record_cassette.py workflow
• preserve structured Responses input, including image content parts, in shared cassette replay helpers
• harden validation, error reporting, storage-disabled behavior, and compatibility with the current request schema

Test Plan

• cargo test
• cargo clippy --all-targets -- -D warnings
• cargo fmt -- --check
• uv tool run pre-commit run --all-files
• smoke-tested record_cassette.py structured-input loading against tests/cassettes/compaction/inputs/basic.json
• verified the four OpenAI compaction cassette replay scenarios within the full Rust test suite

[maralbahari]

@franciscojavierarceo Thanks for the work. I have added some inline comment one is to clarify the auto compaction logic and another comment is for incorrect error type mapping.