Seamless support for multiple backends

[fandreuz]

Support multiple backends (NumPy, PyTorch, JAX, CuPy, ...) by decoupling the implementation of DMD variant from the provider of linear algebra operations. This would enable running on distributed architectures, GPUs, and maybe even differentiation.

This could be achieved by inferring the backend from the type of vectors/matrices provided by the user, and writing some mapping class for functions and methods.

Example (high level idea)

from pydmd.la_mappings import svd

X = np.ndarray
svd(X)

def svd(X):
   if isinstance(X, np.ndarray):
      return np.linalg.svd
   ...

Fixes #297