ENH: sparse: Speed up _add_sparse for DIA format

[perimosocordiae]

This PR rewrites the _add_sparse logic for DIA-format sparse arrays/matrices to avoid expensive allocations and Python looping.

My local benchmarks show that this is faster, in the easy case where the sparsity pattern doesn't change, but also in the more complex case where the sets of diagonal offsets are non-overlapping.

Generated via:
bench.py

[perimosocordiae]

I ran the asv benchmarks, though notably I had to remove the --quick flag from our dev.py runner first. If running in --quick mode, the results are too noisy to be useful.

Change	Before [8ce393f]	After [8eb2614]	Ratio	Benchmark (Parameter)
-	811±50μs	265±40μs	0.33	sparse.Arithmetic.time_arithmetic('dia', 'AA', 'add')

In these benchmarks the A matrix is in DIA format, and B is in CSR format. So the AB, BA, and BB tests aren't exercising the updated code path, which is why they don't show up in the timing diffs.

[dschult]

The code changes here look good!
[I especially like _invert_index which I've done for lists and dicts, but not seen before for numpy.]

The bench.py file checks all-overlap and no-overlap, so I tried some-overlap. Most of these cases were also faster than the main branch. But in one case , the new code is slightly slower than the main branch. I don't believe it is significant enough to keep the old methods. But it might be worth trying the suggestion to check both self and other for whether they include all new offsets.

The above makes B a 3-diagonal poisson1d_in2d constructor with offsets [-1, 0, 1]. Thus they are a subset of the offsets for the poisson2d A.

Most of the variants I tried end up looking like the following, with the new code faster than the main branch. These have B being a poisson2d with offsets [-N, -2, 0, 2, N]. And the same A and B as the some-overlap case above, only B+A instead of A+B. That is what made me think of symmetrizing the new code -- but the main branch code is not symmetric anyway -- B+A is more than 50% different from A+B sometimes.

[perimosocordiae]

Thanks for pointing out the opportunity for the other direction's overlapping-offsets fast path, @dschult. I ran a before and after using a similar setup to what you described, where c's offsets are a subset of a's offsets:

Before 1eba3cb:

After 1eba3cb:

[dschult]

This looks good to me!

[rgommers]

Looks like a nice improvement! Thanks @perimosocordiae, and thanks for the review @dschult

[tylerjereddy]

drafted the release note