fix(acp): route <think> reasoning content to agent_thought_chunk

[alpibrupa]

Summary

• session/prompt streamed every text delta as agent_message_chunk, so a provider's <think>...</think> reasoning content leaked into the assistant's visible message instead of the ACP schema's dedicated agent_thought_chunk update kind.
• Adds per-session in-think state (file-based at /tmp/lex-acp-think-<session_id>, since run_turn_streaming_with_provider's callback type is fixed at (d.Step) -> [io] Unit shared across mcp_main.lex/TUI/client_protocol.lex and can't be widened to [io, sql] for just this caller) and a recursive splitter (split_think) that segments each streamed chunk across <think>/</think> boundaries, routing Thought segments to agent_thought_chunk and Msg segments to agent_message_chunk.
• Also strips a duplicated leading <think> marker — some providers double-emit the opening tag — caught via live testing against a real reasoning model.

Fixes #107

Test plan

• lex check src/server/client_protocol.lex — ok
• lex fmt --check src/server/client_protocol.lex — formatted
• Full repo sweep: lex check over every tracked .lex file — all pass
• lex test — 4/4 passed
• Unit-level: 8 hand-traced split_think cases (plain text, close-then-open, tag spanning a chunk boundary, open-with-no-close, close-with-no-open, doubled leading <think>, and the exact literal <think>The boundary chunk from the original bug report) — all correct
• Live ACP client test (real initialize→session/new→session/prompt→session/close flow against a local reasoning model over litellm): before the fix, 100% of streamed text — including reasoning — arrived as agent_message_chunk; after the fix, reasoning arrives as genuine agent_thought_chunk frames and zero <think>/</think> substrings leak into any chunk's text across a 244-frame run