fix error where model always thinking

[thezukiru]

Summary

• Fix model suffix handling so -thinking and -search explicitly control runtime features.
• Prevent base model defaults from forcing thinking/search mode on non-suffixed model names.

Why

• Previously, if a base model mapping had EnableThinking: true, using a plain model name like GLM-5.1 could still enable thinking mode unexpectedly.
• Users expect GLM-5.1 to run without thinking, while GLM-5.1-thinking should explicitly enable thinking.

Changes

• Updated GetModelMapping so parsed suffix flags override the base model mapping feature flags.
• EnableThinking, WebSearch, and AutoWebSearch are now derived from the requested model suffixes.
• This makes model behavior predictable:
  • GLM-5.1 → thinking/search disabled
  • GLM-5.1-thinking → thinking enabled
  • GLM-5.1-search → search enabled
  • GLM-5.1-thinking-search → thinking and search enabled

Test Plan

• 本地构建通过
• 关键功能已自测
• 如涉及接口变更，已验证兼容性
• 如涉及配置 / workflow 变更，已验证触发逻辑

Checklist

• 已关联相关 Issue（如有）
• 已避免引入无关改动
• 已更新必要文档或说明（如适用）