add a basic implementation of a recording based tool and client

[kshmir]

Some stuff is still quite irrelevant as I was trying out the repo...

1. Made a SnapshotBrowserTool that records the steps taken by the AI.
2. Made a SnapshotAIClient that makes a final decision after all the recorded steps are rerun from the snapshot file (JSONL)

I need to clean it up, would be useful to ignore the screenshot steps, and maybe add a wait for load after clicking or navigating instead of a 1s sleep.

Also a --snapshot parameter should be added and a cleanup parameter as well.

Speed is clearly better, and the AI still makes the final judgement call with a different prompt.