refactor: Experimental optimizations

[crueter]

Some experimental and presently untested optimizations to a few of the scripts. Most of this is bypassing Pydantic's incredibly slow model parsing.

• Larger chunk sizes for download writes, which should reduce syscall overload.
• Avoid pydantic's slow parse_obj method, might be faster?
• Rewrote Neoforge generation to write a dict directly rather than deal with Pydantic nonsense
• Got rid of the thousands of sort operations on recommended_versions
• Faster dict conversion to sidestep Pydantic's recursive parsing

Python sucks.