Support ACP audio prompts

[shreyas-shinde]

Summary

• advertise ACP audio prompt support
• map ACP base64 audio blocks to Codex App Server audio data URLs for normal prompts and steering
• reject audio explicitly when the selected Codex model does not advertise the audio input modality

Why

The adapter currently returns null for every ACP audio block, silently removing user input. Codex App Server already accepts UserInput::Audio data URLs, and model metadata exposes whether audio input is supported. This change preserves supported audio end to end while preventing unsupported models from replacing it with an omission placeholder.

Verification

• 37 test files passed, 331 tests passed
• npm run typecheck
• npm run build
• real authenticated text ACP smoke test completed with end_turn
• a real ACP audio probe reached App Server, which replaced it with an unsupported-audio omission for the selected model
• live model catalog check on Codex 0.145.0 confirmed all 7 current subscription models advertise only text/image, so the new unsupported-model guard is required today

[shreyas-shinde]

Maintainer action needed: this PR is mergeable, but its GitHub Actions run is still blocked as action_required because the workflow has not been approved for this fork. Could a maintainer approve the run and review the ACP audio-content pass-through change? The implementation preserves negotiated capabilities and only forwards audio when the selected runtime advertises audio support; unsupported audio is rejected explicitly rather than silently dropped.