claude_code provider: end-of-turn boundary moved from intermediate
`assistant` snapshots to `stream_event { message_delta }`, fixing tool
loops that were terminating before tool_use blocks were emitted (typical
when extended thinking ran first). Final usage is now taken from
`message_delta`. Thinking blocks (incl. signature) round-trip via the
`anthropic_content` provider_data envelope so reasoning persists across
`claude -p` invocations.

mimo: split into raw::mimo with spec-correct request shape — thinking
type only emits enabled/disabled (no adaptive), output_config dropped,
max_tokens optional, cache_control dropped, repetition_truncation maps
to FinishReason::Length. Routing in request.rs swapped accordingly.
Old stream_mimo / complete_mimo wrappers removed from raw::anthropic.

Tests: 11 new mimo unit tests, new tests/smoke_mimo.rs (3 ignored
live tests covering complete / stream+thinking / multi-turn tool).