Optimise uses of ndarray

[axelboc]

• Split useMemo in two in VisProvider (one for ndarray creation, one for slicing).
• Simplify unpack(mappedView).flat() by unpacking a flat array directly.
• Find domain with cwise.
• Create data points for line vis with cwise.
• Create typed array for heatmap texture without calling flat (pick => shape 1D => unpack), or even without calling Float64Array.from (cwise + push to buffer). - no longer calling flat

[loichuder]

WIP at #178.
Waiting for cwise to support recent ES versions to go further (scijs/cwise#25).

[axelboc]

I think #198 brings us to a good state. The cwise optimisations no longer seem relevant anyway. There are still a few places where the code could be optimised (e.g. findDomain called twice with log scale and negative/positive domain, two uses of findDomain(values.filter()) which amounts to two loops, etc.), but I don't think the gain would be worthwhile when compared to the increase in code complexity.

[loichuder]

There are still a few places where the code could be optimised (e.g. findDomain called twice with log scale and negative/positive domain, two uses of findDomain(values.filter()) which amounts to two loops, etc.), but I don't think the gain would be worthwhile when compared to the increase in code complexity.

Yes, I think the performance is not an issue right now: things are running quite smoothly. Let's wait until we get more complains on that 😄