Support large llm request message input

[cbusillo]

Summary

• add --message-file <PATH> for code llm request
• support --message - to read prompt content from stdin
• require exactly one message source and keep structured request behavior on the existing non-agent ModelClient::stream path
• avoid duplicating streamed output with final OutputItemDone, preserving clean structured stdout

Fixes #336

Validation

• cargo test --manifest-path code-rs/Cargo.toml -p code-cli llm::tests -- --nocapture
• ./build-fast.sh
• JetBrains changed-file inspection: clean
• Read-only agent reviews: GPT-5.5 green; Antigravity green after noting/fixing the pre-existing duplicate-output issue