fix: ensure message history starts with a user turn for Gemini compatibility

[nldhuyen0047]

Summary

Fixes a crash/rejection from the Gemini API (INVALID_ARGUMENT) that occurs when the message history starts with a model (assistant) turn containing a functionCall block.

Root cause

After context trimming (validate_and_repair), the drain boundary can land on an assistant turn, leaving it at position 0. Gemini strictly requires:

• The first message must be a user turn
• A functionCall turn must be preceded by a user turn (not just followed by a functionResponse)

The existing sanitize_gemini_turns only checked that a functionCall was followed by a functionResponse — it did not guard against i == 0 (no preceding user turn).

Fix

1. session_repair.rs — new ensure_starts_with_user(): drops leading assistant turns, re-runs validate_and_repair in a loop to clean up newly-orphaned ToolResult blocks.

2. agent_loop.rs — calls ensure_starts_with_user() after context trim and after overflow recovery (both streaming and non-streaming loops).

3. drivers/gemini.rs — sanitize_gemini_turns now drops functionCall parts when i == 0 as a last-resort defensive guard.

4. channels/line.rs — minor: remove unnecessary borrow on base64::encode call (clippy lint).

Test plan

• Send a message with Gemini after a long conversation that triggers context trimming — should no longer return INVALID_ARGUMENT
• Verify other providers (OpenAI, Groq, Anthropic) are unaffected
• cargo test --workspace passes