Support for out-of-source quantizers

[Giuseppe5]

The goal of this PR is to allow any user to easily change how quantization is applied during the fine tuning process.
A longer description of the issue can be found in #3521

By making _prepare_for_qat a staticmethod, it is possible to replace it like so:

def my_quant_func(...):
    ...
    return model

FastLlamaModel._prepare_for_qat = my_quant_func

As mentioned in the issue above, even allowing for a custom quantization function (and thus, custom quantizer like Brevitas), does not solve all the issues, since Unsloth makes a strong assumption about the names of the quantization functions for weights and activations.

This can be patched without necessarily changing anything within Unsloth, but I believe it might be worth thinking about a more general implementation for the naming conventions (either a task for this or another PR, you tell me).

cc @Datta0