Honor SDK retry configuration

[hi-michael-li]

Why

A second retry layer could multiply provider attempts and make the configured SDK retry limit ineffective.

What

• Let the OpenAI SDK own retries for OpenAI-compatible clients.
• Pass the configured retry limit through once.
• Preserve LiteLLM's existing retry behavior.

Testing

• Focused SDK retry coverage: 2 passed.
• Full suite: 165 passed.
• Ruff, formatting, and diff checks pass. Type checking reports only the existing Slack optional-API warning.

[hi-michael-li]

Planned simplification

I’m going to rebuild this PR directly on main so it contains only the SDK retry-ownership change:

• two OpenAI client files;
• one focused behavioral test module;
• no output-path or web-fetch ancestry.

The production change remains deliberately small: remove Stirrup’s outer retry layer and let the configured SDK retry policy own attempts.

[hi-michael-li]

Simplification complete

• Head: d79a066
• Scope: 24 files (+2517/-483) → 3 files (+52/-48)
• Rebuilt directly on main; removed all output-path and web-fetch ancestry.
• The PR now contains only the two SDK clients and one behavioral retry test module.
• Validation: 2 focused tests; 165 full-suite tests; Ruff, formatting, type and diff checks pass (apart from the existing Slack type warning).