feat: support model alias

[sigoden]

This PR allows the use of name as an alias and real_name as the model value to send to the LLM provider.

Benifits

1. Support the same models with different configuration sets

# o3-mini with different `reasoning_effort`
models:
- name: o3-mini-high  # patch body with extra `reasoning_effort: high`
  real_name: o3-mini
- name: o3-mini-low   # patch body with extra `reasoning_effort: low`
  real_name: o3-mini

# gemini-2.0-flash with search capability
models:
- name: gemini-2.0-flash-online # patch body with `"tools": [{"google_search_retrieval": {}}]`
  real_name: gemini-2.0-flash

2. Support models that only have randomly generated names (#1147)

models:
- name: doubao-1.5-pro
  real_name: ep-xxxx-yyyy

3. Support alias

# Get rid of api version
models:
- name: gemini-flash
  real_name: gemini-2.0-flash-001

# Simplify long name
models:
- name: llama-3.3-70b
  real_name: accounts/fireworks/models/llama-v3p3-70b-instruct