Improve adherence to Python array API Standards #740
Description
Since v1.12.0, PyLops started to support alternative backends to NumPy, with the main goal of being able to perform computations (i.e., apply operators and solvers) on GPUs.
Initially, inter-operating different array/tensor libraries required "a lot" of manual work due to inconsistent APIs and lack of a unifying layer. As such we created the backend module to encapsulate as much of the boilerplate code and limit changes in the actual operator/solver implementations. This allow allowed us to have protected imports of optional libraries and fallback to NumPy when such libraries where not installed in the system.
Over time, the Python ecosystem has caught up with initiatives such as the Python Array API Standards that creates a unified standard that all array/tensor libraries have to adhere to. If a library adheres to the standard, it becomes easier to write a tensor-library agnostic code.
More recently application libraries that serve purposes similar to that of PyLops have started to adopt this standard, with scikit-learn being an example https://labs.quansight.org/blog/array-api-scikit-learn-2026. Also utility libraries have started to appear (e.g., array-api-compat which serve very similar purpose to what we do in the backend module.
It is maybe about time to start looking into this and unifying ourselves to the standard that started to appear and is now quite widely recognized and adopted.