refactor: centralize model parameters into JSON config (model_specs)

[Haixing-Hu]

refactor: centralize model parameters into JSON config (model_specs)

Replace hardcoded model-specific parameters (maxOutputTokens, thinkingBudget
caps, overhead values, model aliases) with a centralized JSON config file
(resources/model_specs.json) and a Rust registry module (proxy/model_specs.rs).

New files:
- resources/model_specs.json: all model parameters in one place
- src/proxy/model_specs.rs: config loading, lookup (with prefix matching),
  and convenience functions (cap_max_output_tokens, cap_thinking_budget, etc.)

Refactored files:
- common_utils.rs: model alias resolution via model_specs::resolve_alias()
- gemini/wrapper.rs: default thinking budget, budget capping, maxOutputTokens
  cap, adaptive max tokens, thinking overhead — all from config
- openai/request.rs: default budget, budget caps, overhead values from config
- claude/request.rs: default budget, budget caps, opus fixed values,
  maxOutputTokens safe limit, overhead values — all from config

Benefits:
- Adding new models or adjusting limits requires only editing model_specs.json
- No more scattered magic numbers (24576, 65536, 131072, 16000, etc.)
- Consistent behavior across all three protocol mappers