New rule: use `dtype` instead of converting after creation

[kit1980]

For many ops dtype parameter can be provided, and it should be preferred (from readability and performance point of views) to use it instead of converting the op results to a different type.

For example, torch.arange(1, N).float() should be changed to torch.arange(1, N, dtype=torch.float32)

cc @malfet