add keepzeros option in sparse

[matthieugomez]

issue #12605

[tkelman]

Please squash the commits here. This could use quite a few more tests, and documentation. This should probably be benchmarked as well to check whether the addition of a keyword argument makes a noticeable difference to the performance of sparse. There are some packages, Convex.jl as one example, that spend quite a bit of time calling sparse in inner loops.

[ViralBShah]

I would like to hold this till 0.5, and not want to do it this late.

There is the obvious issue here that the moment you do some other sparse matrix operation that uses this matrix, the zeros will get squeezed out in the result. Once 0.4 is done, I would like to revisit the design issues with stored zeros.

[matthieugomez]

I'm perfectly fine with manipulating internals using the functions rowvals zrange and nonzeros afterwards. I then use some matrix multiplication / division. All these operations keep the structural zeros right?

[ViralBShah]

Not at the moment.

[mlubin]

This is a really important and useful feature to have, but I agree that it's past the window for getting into 0.4.

[ViralBShah]

@tkelman What do you think we should do here?

[tkelman]

Gradually move away from matlab's approach to stored zeros and more towards a scipy design. A keyword argument might not be the best way to accomplish this though.

[ViralBShah]

Ok - perhaps the hybrid is not a great approach. Let's just do that slowly and hopefully by the 0.5 release.

[tkelman]

#14798 covers this and a bit more, so closing. Sorry this didn't go in as-is @matthieugomez.

[matthieugomez]

Thanks!