feat(kernel): expose reasoning trace on blocking InferenceResponse

[amos-aios]

Summary

InferenceResponse (the blocking-path result) had no field for a model's
deep-reasoning trace, even though the streaming path surfaces one via
StreamChunk::Reasoning. Reasoning-capable models (DeepSeek-R1 style) lose
that trace entirely for non-streaming callers.

This adds an optional, serde-defaulted reasoning: Option<String> to
InferenceResponse, mirroring the streaming representation:

/// Deep-reasoning trace surfaced by reasoning-capable models on the
/// blocking path (`None` otherwise). Mirrors what the streaming path
/// delivers as `StreamChunk::Reasoning` increments.
#[serde(default, skip_serializing_if = "Option::is_none")]
pub reasoning: Option<String>,

• #[serde(default)] + existing ..Default::default() update syntax means
  zero changes are required at any backend construction site.
• Wire-compatible in both directions: absent on ordinary completions,
  skipped on serialize when None.

Motivation

Found while embedding the engine as an in-process library
(mofa-studio): the OpenAI-compatible gateway there forwards the blocking
reasoning_content field for R1-style traces and round-trips engine
responses through typed structs, so the missing field silently dropped the
trace. This restores parity between streaming and blocking paths.

Testing

• cargo test -p mofa-kernel — 13 passed.
• Full workspace cargo check/test green with the field added; no call-site
  churn anywhere.