Add user-controlled reasoning depth selection (Light/Medium/Deep/Auto) to let users balance speed vs. thoroughness based on their task needs.

## Summary

Add user-controlled reasoning depth selection (Light/Medium/Deep/Auto) to let users balance speed vs. thoroughness based on their task needs.

## Component(s) Affected

- [x] Core / Backend (app, agents, config, providers, utils, local_models)
- [x] Console (frontend web UI)
- [ ] Channels (DingTalk, Feishu, QQ, Discord, iMessage, etc.)
- [ ] Skills
- [x] CLI
- [ ] Documentation (website)
- [ ] Tests
- [ ] CI/CD
- [ ] Scripts / Deploy

## Problem / Motivation

Users have different needs for reasoning complexity:

- **Quick tasks**: Simple questions need fast, direct answers without extended reasoning
- **Complex analysis**: Deep problems require thorough thinking, multiple perspectives, and verification
- **Token budget**: Extended reasoning consumes more tokens; users should control this tradeoff
- **Time sensitivity**: Some situations prioritize speed over thoroughness

Currently:
- No standardized way to control reasoning effort
- Users can verbally request "keep it simple" or "think deeply," but this isn't enforced
- No visual indication of reasoning mode
- Token usage varies unpredictably

## Proposed Solution

1. **UI Control**: Add a dropdown or slider in Console to select reasoning depth:
- **Light**: Fast, minimal reasoning, direct answers (low token usage)
- **Medium**: Balanced reasoning (default)
- **Deep**: Extended reasoning, thorough analysis, multiple perspectives (high token usage)
- **Auto**: Agent decides based on task complexity

2. **Backend Implementation**:
- Adjust model parameters (max_tokens, temperature, etc.)
- Control reasoning steps or chain-of-thought length
- Optionally route to different model tiers

3. **Visual Feedback**:
- Show current reasoning mode in chat header
- Display token budget estimate
- Indicate when mode affects response time

4. **CLI Support**: Add flag or slash command (e.g., `/reasoning deep`)

## Alternatives Considered

- **Verbal instructions**: "Think carefully about this" — inconsistent and not enforced
- **Model selection**: Choose different models manually, but this doesn't control reasoning depth within a model
- **Token limits**: Set max_tokens globally, but this cuts off responses rather than controlling reasoning style

## Additional Context

- Similar features exist in other AI platforms (e.g., "effort" settings, "quick vs. detailed" modes)
- Could integrate with QwenPaw's existing model selection system
- May affect: token usage, response time, answer quality, tool usage patterns
- Relevant for both cloud and local model deployments

## Willing to Contribute

- [ ] I am willing to open a PR for this feature (after discussion).