[codex] Add vision transcriber support

[cannebergeee]

Summary

• Add visionTranscriber settings and DEEPCODE_VISION_* / VISION_* resolution for non-multimodal image handling.
• Add a vision transcription module that sends image parts to an OpenAI-compatible vision model and inserts hidden transcript messages.
• Wire session activation and message conversion so multimodal models keep original images, while text-only models use transcripts and fail explicitly when transcription is unavailable.
• Update README/configuration docs and add settings, converter, and session coverage.

Validation

• npm run check passed.
• npm run test:single -- src/tests/settings-and-notify.test.ts src/tests/openai-message-converter.test.ts passed: 52 pass, 1 skip.
• npm run test:single -- src/tests/session.test.ts passed: 88 pass, 5 skip.
• npm test completed with one pre-existing/unrelated failure in Bash terminates commands that exceed the configured timeout; it reproduces in isolation and this PR does not touch Bash handler code.