Break out inner loop of optimize_layout so joblib can parallelize it (0.4dev version)

[tomwhite]

This is the 0.4dev version of #255. Probably best to review this one.

If no random seed is set then optimize_layout will use all cores on the machine, otherwise it will use a single core for reproducibility.

[pep8speaks]

Hello @tomwhite! Thanks for opening this PR. We checked the lines you've touched for PEP 8 issues, and found:

• In the file umap/layouts.py:

Line 262:80: E501 line too long (86 > 79 characters)
Line 299:80: E501 line too long (116 > 79 characters)

• In the file umap/umap_.py:

Line 1892:80: E501 line too long (99 > 79 characters)

[sleighsoft]

I am not super deep into what you are doing here, so bare with me if my question makes no sense.
Why can't you use @numba.njit(parallel=True) @ line 129 instead of using nogil=True and joblib?

[tomwhite]

Your question makes a lot of sense. The answer is to do with random number seeding and reproducibility: basically Numba parallel doesn't allow controlled seeding across threads so it's not possible to get a reproducible result. With this code (using joblib), if you don't set a seed it will use all cores, otherwise it will use a single core and the result will be deterministic. See discussion in #231 for why that is not achievable with Numba parallel.

[sleighsoft]

You could just toggle the decorator.

import numba
import numpy as np

def searchsorted(a, v):
    indices = np.empty(v.shape, dtype=np.int64)
    for i in numba.prange(v.shape[0]):
        indices[i] = np.searchsorted(a[i], v[i])
    return indices

# No numba, no parallel
searchsorted(np.array([[4,3,2,5]]), np.array([[1]]))

# Now uses the numba implementation
parallel_searchsorted = numba.njit(searchsorted)

parallel_searchsorted(np.array([[4,3,2,5]]), np.array([[1]]))

[sleighsoft]

This way you can also programmatically just toggle parallel on or off by passing it to njit.
See https://github.com/numba/numba/blob/master/numba/decorators.py#L41

[tomwhite]

@sleighsoft you are right, this would be a better approach, since it's shorter and doesn't need a new dependency (joblib). I actually investigated this approach a while ago, but it got lost in the discussion in #231. I think the key point (which I missed before) was that we can use numba parallel if no seed is set (to get speed), but disable parallel if a seed is set (to get reproducibility).

I've opened #294 with the suggested approach.

[tomwhite]

Closing in favour of #294