Improve DBSCAN performance

[jonatanklosko]

Closes #324.

For the reproduction from #324 (comment):

before	after
0.140113 s	0.148038 s
0.08597 s	0.070765 s
0.663867 s	0.083935 s
16.370221 s	0.084186 s
109.480284 s	0.087279 s
381.106067 s	0.1175 s
	0.147327 s
	0.174654 s

[jonatanklosko]

The max stack size was heavily overestimated. The stack is effectively a set of points and should never exceed N points. We just need to make sure to never add duplicates into the stack, which the other line change does.

[jonatanklosko]

My first attempt was to remove the innermost while loop (putting all relevant indices onto the stack at once), and I was able to with a trick. But I started to think that it may break in an edge case if we get to a close stack. Then analyzing the stack size, I realised it should be way smaller, and that was the actual fix :p

[krstopro]

Looks good to me.

One thing we might want is more unit tests for these algorithms. For example, ten points at one location (duplicates) and ten points at another one far away. I could do this as I expect to have some time at the end of the year.

[josevalim]

Awesome job!

[jonatanklosko]

I came up with a rewrite that paralelizes much better, I am merging this for reference, but will submit another PR soon :)

[jonatanklosko]

One thing we might want is more unit tests for these algorithms. For example, ten points at one location (duplicates) and ten points at another one far away. I could do this as I expect to have some time at the end of the year.

Definitely, PRs for that would be great!