add snapshot tests

[DhruvDh]

This draft PR adds snapshot tests. I am trying to incrementally rewrite operations in KernelAbstractions.jl, and am using these tests to ensure consistent outputs at temp 0.

Currently, it doesn't run on GPUs, for that I think a method that moves weights to a chosen backend would be needed.
Right now, I just want to write correct kernels that run on the CPU and produce outputs consistent to the original version.

Here is a kernel for rmsnorm - DhruvDh@1d49d42. It makes the whole thing some 2-5% slower but produces correct outputs. This way, we can try incrementally writing kernels for operations and later look at the results and decide if GPU acceleration is needed, how to maintain readability, etc.

I can also keep a draft PR open where we do this incremental addition of kernels

I am also having trouble with formatting, if a .JuliaFormatter.toml can be added to the project, that would help a lot.

[cafaxo]

GPU support would be awesome. I am not sure if testing for exact reproduction is that useful since language models are very sensitive to rounding/quantization errors.
Even llama.cpp produces different outputs depending on the backend (see ggerganov/llama.cpp#4755).