add support for nn.Linear layers in StandardParametrizator

[sapiosaturn]

Hi,

I was training a small model using nn.Linear layers, since I had no immediate need for tensor parallel. This resulted in the following error (truncated):

[rank0]:   File "/root/nanotron/src/nanotron/scaling/parametrization.py", line 28, in parametrize
[rank0]:     raise Exception(f"Parameter {param_name} was not initialized")
[rank0]: Exception: Parameter weight was not initialized

I looked at the file and realized that the parameter initialization logic didn't have support for nn.Linear layers, so I added it myself. I thought I'd make a small PR in case anyone else runs into the same issue when just using nn.Linear layers at small/medium scale.