Support GGUF version 2 as well as 3.

[ScottTodd]

Version 3 seems to have only added support for big endian data, which I don't think we care about? Other changes have been made without updating the version (like adding bf16 support), so we may need to update this code with other logic.

~~Not sure I like the testing changes I made here (parameterized tests). Could do something simpler, especially if we want to only test things like bf16 on newer versions.~~