Handle 413 prompt too large response from APIs like vertex gracefully

[ianlintner]

As a user if roo detects a prompt too long response rather than getting stuck in a retry loop could handle it automatically without user input.

e.g. From google vertex ai

413 {"type":"error","error":{"type":"invalid_request_error","message":"Prompt is too long"}}

[ericfitz]

+1

It's not just vertex AI (Google's models have a 1M token window, one of the largest). It's worse on Anthropic's models that have a 200K token window - trying to read one nontrivial npm lock file will blow that away every time. I have specifically changed my prompts for one project to exclude lockfiles for this reason.