fix: parse Ollama reasoning fields

[code-yeongyu]

Summary

• parse Ollama/Qwen reasoning fields in OpenAI-compatible non-streaming and streaming responses
• add focused integration coverage for reasoning-before-text ordering
• hint local Ollama OPENAI_BASE_URL for tagged Qwen model names and document local setup

Validation

• cargo fmt --all --check
• cargo test -p api ollama_reasoning -- --nocapture
• cargo test -p api reasoning_content -- --nocapture
• cargo test -p api openai_compatible_client_honors_http_proxy_for_requests -- --nocapture
• cargo test -p api local_openai_gateway_strips_routing_prefix_and_preserves_extra_body_params -- --nocapture
• cargo test -p rusty-claude-cli qwen_invalid_model_hint_mentions_local_ollama_openai_base_url -- --nocapture
• cargo test -p rusty-claude-cli parses_single_word_command_aliases_without_falling_back_to_prompt_mode -- --nocapture
• cargo test --workspace
• cargo clippy --workspace --all-targets
• Manual QA: OLLAMA_HOST=http://127.0.0.1:11434 ./rust/target/debug/claw --model gpt-oss:20b prompt "Reply exactly HELLO_WORLD_123" returned HELLO_WORLD_123

Notes

• Qwen-specific live QA was not run because no Qwen/Qwen3 Ollama model is installed locally; installed models are gpt-oss:20b and kimi-k2.6:cloud.
• PR fix(providers): preserve OpenAI-compatible reasoning history #3234 routing/model preservation regressions were checked with focused proxy/local gateway tests.

[1716775457damn]

This is a necessary companion to #3228. Parsing the Ollama reasoning field properly ensures Qwen3 thinking models produce usable output instead of empty responses. The CI is green and the fix is clean - good to go.