feat(autofix): Use an LLM to apply edits to full files in the coding step

[roaga]

Inspired by Cursor's apply model:

• remove strict diff requirements from the coding prompts
• apply unstructured code suggestions to full files with an LLM (gpt-4o-mini) using speculative decoding (300-500 tokens per second from my observations)
• automatically create the diff using unifieddifflib
• process code changes in parallel

Pros:

• improved fix evals by 4 points up to 41% (after all these changes, we're up to 88% correct RC and 41% correct fix on the new evals)
• removes unreliability of incorrect diffs, missing changes, and poorly applied changes (which should solve some of the issues that have been plaguing us forever, including recent reports)
• usually insignificant difference in overall latency. It takes a few seconds to apply, but reduces tokens the main agent needs to write and removes incorrect diff fixer.

Cons:

• adds more latency if editing a very big file (in the future, we could consider using a smaller model on Fireworks with predicted output for even more speed; that's what Cursor does)

[roaga]

this fixes the bug we've had for a while where it adds random empty lines in the diff