Add nnz parameter to sparse.random

[emilmelnikov]

Currently, sparse.random uses the density argument to determine
the number of non-zero elements in the generated array. If the exact
number of non-zero elements is required, one might try to compute
density as nnz / np.prod(shape), which might give a wrong result
for large values due to a floating-point error.

This PR adds a new, optional nnz argument to sparse.random that,
if specified, overrides density and allows a caller to create
a random array with the exact number of non-zero elements.

[codecov]

Codecov Report

Merging #410 into master will increase coverage by 0.01%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master     #410      +/-   ##
==========================================
+ Coverage   95.05%   95.06%   +0.01%     
==========================================
  Files          19       19              
  Lines        2810     2819       +9     
==========================================
+ Hits         2671     2680       +9     
  Misses        139      139              

[hameerabbasi]

Hello. This is a nice addition, I'd just like to suggest tests for:

• Err on something outside the unit interval.
• Err when both density and nnz are provided.

Use pytest.raises (there are a couple of usages in this file).

[hameerabbasi]

I think there was a misunderstanding about what I meant: If exceptions are raised, they fail the test anyway (which is what we want). I meant you should test that:

1. The exceptions you're raising are actually raised (use pytest.raises, grep for it in the codebase).
2. Change the default density to None.
3. Raise an error if both nnz and density are not None.
4. If both are None, assume a density of 0.01.

[hameerabbasi]

One minor change more, otherwise LGTM.

[hameerabbasi]

Thanks, @emilmelnikov!