perf(reasoning): dedupe to_lowercase + marker early-outs in Qwen apply_reasoning_mode (#402)

[kiannidev]

Problem

Every Qwen chat/voice utterance runs apply_reasoning_mode before the LLM sees messages. On main, explicit_reasoning_mode, is_simple_request, and looks_like_deep_reasoning_request each call to_lowercase() independently — up to three full-string allocations on the same user text.

Proposed fix

Behavior-preserving optimization in crates/genie-core/src/reasoning.rs:

• Single shared to_lowercase() in the Qwen path.
• Conservative needs_simple_request_scan / needs_deep_reasoning_scan early-outs before expensive marker loops.
• Corpus regression tests + optional release bench.

Scope

Pre-LLM performance bucket under #402. Does not touch tools/quick.rs (avoids collision with open quick-router PRs). Complements #501 (broader pre-LLM bundle) and merged voice-intent work.

Acceptance

• ReasoningDecision + adjusted message content byte-identical to main for a regression corpus.
• cargo test -p genie-core + cargo clippy + cargo fmt clean.