feat: add backoff_factor and max_backoff for customizable retry backoff

[crawfordxx]

Summary

Adds two new optional parameters to the OpenAI and AsyncOpenAI client constructors:

• backoff_factor (float | None) — multiplier for the exponential backoff delay between retries (default: 0.5 seconds, matching current behavior)
• max_backoff (float | None) — maximum wait time between retries (default: 8.0 seconds, matching current behavior)

These parameters are threaded through BaseClient, SyncAPIClient, AsyncAPIClient, and the copy()/with_options() methods so they work consistently everywhere.

Usage

# Default behavior (unchanged)
client = OpenAI()

# Longer backoff for rate-limited workloads
client = OpenAI(
    max_retries=5,
    backoff_factor=2.0,   # start at 2s instead of 0.5s
    max_backoff=60.0,     # allow up to 60s between retries
)

# Also works with with_options/copy
slow_client = client.with_options(backoff_factor=5.0, max_backoff=120.0)

Backward compatibility

When backoff_factor and max_backoff are not provided (the default), the existing constants INITIAL_RETRY_DELAY (0.5) and MAX_RETRY_DELAY (8.0) are used, so there is zero behavior change for existing users.

Closes #1102

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: d5d2ca507a

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

[chatgpt-codex-connector]

Avoid forwarding backoff kwargs into incompatible subclasses

OpenAI.copy() now always forwards backoff_factor/max_backoff into self.__class__(...), which breaks subclasses whose __init__ signatures do not accept those keywords. In this repo, AzureOpenAI.copy() and AsyncAzureOpenAI.copy() delegate to super().copy(), so calling copy()/with_options() on either Azure client now raises TypeError for unexpected keyword arguments, regressing an existing workflow for Azure users.

Useful? React with 👍 / 👎.