Reduce 4.1 usage's of insert_edit_into_file

[connor4312]

In swebench we saw 4.1 loves apply_patch and rarely uses insert_edit_into_file. However, in the real world, our telemetry indicates that it still calls it more than we would like. Some ideas:

1. See if we'd be able to remove insert_edit from 4.1's capabilities
2. Or, keep it in with instructions like "Never use this tool unless told otherwise" and have apply_patch in its output say "You may use insert_edit for this edit" only if apply_patch fails twice in a row.

cc @isidorn @roblourens

[isidorn]

I just checked GPT 4.1 telemetry with pre-release extension (insiders users), and it looks like apply_patch is one of the most popular tools with it being used 30% of all too calls. insert_edit_into_file is used around 1.5%. So around ~20x times less.
@connor4312 if you see more - make sure you only look at data with latest stable / pre-release. Since users on older versions are not affected by our experiment.

Option 1 sounds good. We can do the same use_apply_patch_exclusively experiment as we did for Sonnet 4.