Add Cerebras rate limit handler

[kevint-cerebras]

Use exponential backoff to improve UX with Cerebras models.

Since the hourly and daily rate limits for tokens are the same, TPM is the limiter -> max retry wait time = 60 seconds so users can fully take advantage of using Cerebras

[eloquence]

I've tested this PR with Cerebras and can confirm that the user experience is much improved - requests do get throttled, but you no longer get a timeout pretty much immediately.

[JC1738]

Same, I tested on my fork and made using Cerebras doable.