fix(openai): support max_completion_tokens for reasoning models

[maxchang3]

Fixes raycast/extensions#26761

Background

According to the OpenAI docs, reasoning models (e.g. o1, o3, gpt-5) require max_completion_tokens, and max_tokens is not supported:

Currently, our extension uses max_tokens for all models, which leads to incompatibility with reasoning models.

For OpenAI-compatible providers, behavior is inconsistent:

• Some still expect max_tokens
• Others follow OpenAI’s newer convention (max_completion_tokens)

This creates ambiguity when integrating with different providers.

---

Changes

• Introduced model-based detection for reasoning models using pattern (From Dify’s approach):

  /^(o1|o3|gpt-5)/i

• Added endpoint-aware logic:

  • When using the official OpenAI endpoint (https://api.openai.com/)
  • AND the model matches a reasoning model
  • Automatically switch to max_completion_tokens

• Added a user-configurable option:

  • forceMaxCompletionTokens
  • Allows forcing max_completion_tokens regardless of endpoint/model
  • Useful for OpenAI-compatible providers that follow the new convention

[maxchang3]

For this PR, please use Squash and merge

[maxchang3]

Although current models seem to only accept max_completion_tokens, I'll keep this regex to support OpenAI-compatible providers or third-party platforms that also expose OpenAI models, in case it's still needed.