Numpy caching of data in online monitor storage

[JoranAngevaare]

What is the problem / what does the code in this PR do
In #346, we uttered the prophetic words
It's just the backend-buffer that is sitting otherwise idle in the background eating your RAM that needs clearing.
We now learned that this is true when we have longer than usual runs and data with large volumes.

We also should have known that the sys.getsizeof wasn't taking the total size of the document but only of the dictionary.

This PR fixes both with using numpy for the buffers such that the max size of buffers is respected by the backend.

[coveralls]

Coverage decreased (-0.1%) to 91.678% when pulling 47e5894 on numpy_mongo_buffers into 1562542 on master.

[JoranAngevaare]

For bookkeeping, this is about 10x faster:

n=int(1e4)
bla = list(range(n))
n_bla = np.zeros(len(bla), dtype=[('a', np.int64), ('b', np.int64), ('c', np.int64)])
bla_d = [{'a': i, 'b': i, 'c':1} for i in bla]


%%timeit
for i in range(len(bla)):
    for key in ['a', 'b', 'c']:
        n_bla[i][key] = bla_d[i][key]
22.4 ms ± 97.1 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)


%%timeit
for key in ['a', 'b', 'c']:
    n_bla[key] = [dd[key] for dd in bla_d]
1.98 ms ± 50.1 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)