Performance optimizations, 10x for moving sums

[sairus7]

Though I'm not sure about floating point error accumulation. The same thing could be done for runvar, runsd, runcov, etc.

[dysonance]

@sairus7 Thanks for the PR, these look like pretty solid improvements! Could you try to work on getting the build tests to pass? If those get handled then I wouldn't mind merging.

[sairus7]

That error is because of NaNs at the front of the result array. I can fix it in different ways: 1) Initialize sum with the first point, multiplied by window length n - result will start smoothly from the first point. 2) Compute sum with narrower window at the front (window = 1, 2, 3, ... n ) - result will be more noisy at the front. 3) Repeat first valid sum (1:n) at first n-1 points of the output - there will be constant level at the front.

Also, I can update main loop to safely treat NaNs, e.g. for every NaN it can reset its state and output nans, and then initialize with abovementioned options 1, 2 or 3. If we choose option 2, I can discard any NaNs from the running sum, so it will output running sums over n or less points, and if n' == 0, it will output NaN. (More options here - we can even wish to output NaNs if current n' is less than a given value depending on the window size n.)

What do you think?

[sairus7]

Fixed to be consistent with original version, any single NaN that is encountered within a moving window leads to NaN output. Since the code became less clear, I left the old version in comments.

[dysonance]

@sairus7 Nice! Thanks for working on this, seems like a solid improvement.

[sairus7]

See also my newly published package: https://discourse.julialang.org/t/ann-maxminfilters-jl-fast-streaming-maximum-minimum-within-moving-window/30151