remove dead code from fast_rms_layernorm_inference

[KareemMusleh]

If XX is None it will error out on

XX *= variance.rsqrt_()

also I'd like to know:

1. why X.copy_(XX) over X[:] = XX
2. what is the purposed of doing torch_square = torch.square outside the function?

Thank you for your work on unsloth!

[danielhanchen]

Sorry on the delay - I think technically I removed error msgs for XX = None to reduce CPU overhead - it's not much though.

.copy_ I guess also reduces CPU overhead vs [:] - it's the same though