Support streaming LLM responses for large diffs

[ajianaz]

Description

Large diffs can take a long time to review when waiting for the full LLM response. Implement streaming so results appear incrementally.

Acceptance Criteria

• Stream LLM responses token-by-token to the terminal
• Show review findings as they are generated
• Maintain structured output when streaming (SARIF, JSON)
• Fallback to non-streaming for providers that do not support it
• Add --stream / --no-stream flag

Implementation Notes

• OpenAI and Anthropic both support SSE streaming
• Use tokio streams and futures-util for stream processing
• Buffer partial findings until they are complete before displaying severity

[ajianaz]

Implemented in PR #24, merged to develop.