Handle mixed input dtypes and empty size in numba lapack functions

[ricardoV94]

Still WIP, spinoff from #811

Major changes:

1. Systematic fallback to obj mode for complex inputs
2. Systematic casting of discrete inputs to floats (with a warning on compile_verbose)
3. Systematic upcasting of inputs in operations with more than one input (still WIP)
4. Change view(inp).ctypes -> inp.ctypes
5. Explicitly handle empty inputs, as lapack tends to raise

Re: Change view(inp).ctypes -> inp.ctypes

This may have be needed when working with complex inputs? But we are not supporting them in most implementations, so it makes code more complex and is a potential source of bugs when we fail to systematically upcast input (point 3. from above)

Say we have a float32 and a float64 inputs, and forget to upcast the first one. Calling view will raise for non f-contiguous inputs (which we always need for these routines):

import numpy as np x = np.asfortranarray(np.eye(3, dtype="float64")) x.view(dtype="float32") # ValueError: To change to a dtype of a different size, the last axis must be contiguous

Even if it didn't raise the meaning of the array would be nonsensical:

np.eye(3, dtype="float64").view(dtype="float32") # array([[0. , 1.875, 0. , 0. , 0. , 0. ], # [0. , 0. , 0. , 1.875, 0. , 0. ], # [0. , 0. , 0. , 0. , 0. , 1.875]], dtype=float32)

I think we need to call different routines for complex inputs anyway?

[ricardoV94]

We should redo _check_scipy_linalg_matrix to take a dtype as well, so if an input dtype does not match the operation dtype it raises during numba typing