Support for `memmap` in `Quantity`

Description

I if possible, would like the Quantity object to support memmap arrays, meaning that given a memmap array, a quantity constructed from that array remains memory mapped.

Additional context

It is possible that Quantity supports this, but it remains unclear. Right now if I run:

import numpy as np
import astropy.units as u

a = np.memmap('test', mode="w+", shape=(3,4))
q1 = u.Quantity(a, u.kg)
q2 = u.Quantity(a, u.kg, copy=False)

print(f'{isinstance(a, np.memmap)=}')
print(f'{isinstance(q1, np.memmap)=}')
print(f'{isinstance(q1.value, np.memmap)=}')
print(f'{isinstance(q2, np.memmap)=}')
print(f'{isinstance(q2.value, np.memmap)=}')

It returns

isinstance(a, np.memmap)=True
isinstance(q1, np.memmap)=False
isinstance(q1.value, np.memmap)=False
isinstance(q2, np.memmap)=False
isinstance(q2.value, np.memmap)=False

Which from the memmap documentation indicates to me that the the data is no longer memmap.

Basically, I would like a way to pass a memmap array to a Quantity and that Quantity to use the array data through the memmap (ie the data on disk mapped via the memmap to the array gets updated when I update the Quantity). This includes a way to interrogate the Quantity to confirm it is or is not using a memmap.

[mhvk]

Indeed, not immediately. I briefly tried whether it would work to define a new class QuantityMemmap(u.Quantity, np.memmap) but that in itself did not seem good enough. I have not investigated why, though think in principle it should be possible to get it to work; it might need the kind of automatic class construction that was used for Distribution and Masked and which @nstarman has been wanting to generalize. With that, Quantity might also be able to wrap dask arrays, etc.

[mhvk]

In principle, it should be possible to make this work, since np.memmap is an ndarray subclass, but right now our use of ndarray is perhaps a bit too much built in. The suggestion you got from the documentation does work:

a = np.memmap('test', mode="w+", shape=(3,4), dtype='f8')
# Avoid myself making a mmap (should be quite easy, but just testing)
mm = np.ndarray(buffer=a.base, shape=(3, 4), dtype='f8')
q = u.Quantity(a, u.m, copy=False)

# Check state of a
a
memmap([[0., 0., 0., 0.],
        [0., 0., 0., 0.],
        [0., 0., 0., 0.]])
# Set an element in q
q[-1, -1] = 1*u.cm
# Check whether this propagates to a.
a
memmap([[0.  , 0.  , 0.  , 0.  ],
        [0.  , 0.  , 0.  , 0.  ],
        [0.  , 0.  , 0.  , 0.01]])
# Check it is written to disk by exiting and logging in again
a = np.memmap('test', mode="r", shape=(3,4), dtype='f8')
# result is recovered