Right now StorageImpl records a dtype. We should remove the dtype and store bytes directly. Non-POD dtypes should still support custom destruction, but that is built on top of an otherwise untyped representation. Representing storage impls in an untyped way means that we can reinterpret cast tensors, e.g., view a float tensor as a complex tensor.

In the user side we still expose ByteStorage/LongStorage; these should either be eliminated from Python, or implemented as wrappers that know dtype.

cc @ezyang @gchanan @zou3519 @anjali411 @dylanbespalko