ENH: np.memmap should take a strides arguemnt in the construtor

[alanhdu]

I am trying to share a np.memmap array across two processes by sharing the underlying mmap-ed file. To fully replicate this across processes, I need to replicate the filename (to share the buffer) and the numpy metadata. Unfortunately, the np.memmap constructor does not take a strides argument, which means that I can't directly construct the appropriate array.

I can currently achieve my aim by doing:

memmap = np.memmap(filename, dtype=..., mode="r", offset=...)
memmap = np.lib.stride_tricks.as_strided(memmap, shape=..., strides=..., subok=True)

but I think it'd be nice for np.memmap to directly supported the strides argument.

Implementationally, I think we just need to update the num_bytes calculation to take into acount striding and to pass strides into the ndarray.__new__ call at

numpy/numpy/core/memmap.py

Lines 269 to 270 in a14c412

	self = ndarray.__new__(subtype, shape, dtype=descr, buffer=mm,
	offset=array_offset, order=order)