Make sure we pass device="meta" to TE layers during meta device init

[pstjohn]

TransformerEngine requires that we pass device="meta" (and importantly not device=torch.device("meta")) to layer constructors to initialize parameters on the meta device.

This makes sure we pass the right device to the layer constructor and adds tests to ensure the parameters are actually being placed on the right devices.

For TransformerEngine layers, we want to be moving parameters from the meta device to cuda with

for module in model.modules():
    if hasattr(module, "reset_parameters"):
        module.reset_parameters()

while for HF layers (e.g., nn.Embedding), we want to be doing

model.to_empty("cuda")
model.apply(model._init_weights)

to ensure that we pick up the config.initializer_range initialization correctly.

The issue is that we can't do to_empty("cuda") or _init_weights on TE layers, nor can we do reset_parameters() on the HF layers without the preceeding to_empty, and this doesn't use the HF config when creating initial values.

Closes BIO-1