Stateful implementations of existing functions

[s-cavalier]

Currently, the implementation for the MVP uses pure functional approaches with in/out Tensors as effectively just fancy data storage. To make things convenient, I think going forward we should prefer an stateful approach like PyTorch. For example, instead of calling the raw matmul and raw add from Sachit's BLAS integration, we make a dedicated Linear class that is constructed with the Tensor to the data. We extrapolate this for all of the existing functions.

Action items:

• Create a slew of objects that follow the same interface, which is that they expose a callable operator() with parameters const TensorBase<T1, Ext1>& in, TensorBase<T2, Ext2>& out and void return type. For example, Linear should be callable as Linear<TensorType1, TensorType2>{t1, t2}(t_in, t_out); and TensorType1 should be a rank-2 mat along with TensorType1 as a rank-1 mat, both enforced at comptime.
• Do the same for convolve, batchnorm, maxpool, maybe a special Expression-based operation (not necessary), and anything else that would be convenient and in scope.
• I think it could also be really interesting to consider using non-type template T* parameters since most if not all of our weights are externally linked, although I don't know how this could be done ergonomically. Feel free to look into this or ignore.