Add --num-retries option to control API retry behavior

[potatoqualitee]

Issue

Problem

When API calls fail (due to configuration errors, auth issues, rate limits, etc.), aider uses litellm's default retry behavior which can cause users to wait extended periods (60+ seconds) before getting an error message. This is particularly frustrating when the underlying issue isn't transient and retries won't help.

For example, with Azure configuration errors, users currently experience:

Retrying in 8.0 seconds...
Retrying in 16.0 seconds...
Retrying in 32.0 seconds...

This adds up to nearly 1.5 minutes of waiting before seeing the actual error.

Proposed Solution

Add a --num-retries flag (and corresponding AIDER_NUM_RETRIES environment variable) to allow users to control retry behavior:

aider --num-retries 0  # Disable retries completely
aider --num-retries 1  # Retry once

This value should be passed through to litellm's num_retries parameter in completion calls.

Use Cases

• Development/debugging: When testing API configurations, developers want immediate feedback rather than waiting through retry cycles
• CI/CD environments: Automated systems benefit from fast failures
• Production applications: Some users prefer to handle retries at a higher level in their architecture
• Cost control: Fewer retries = fewer billable API calls when issues occur

Alternative Considered

Users could set LITELLM_NUM_RETRIES=0, but it's unclear if aider respects this environment variable, and an explicit aider flag would be more discoverable and consistent with aider's configuration patterns.

Version and model info

No response