Use LoopVectorization.@turbo for evaluation loops

[MilesCranmer]

This switches from @inbounds @simd to @turbo for most of the inner evaluation loops. There is a 30% speedup which is huge.

However, there's currently a StackOverflowError when running the Float16 tests. I'm also not sure if LoopVectorization will be safe to use, since ideally I want to allow for any user-defined operator - would it mean certain passed operators fail? Also need to test it on a distributed system.

[MilesCranmer]

Okay looks like LoopVectorization.jl hasn't even implemented SpecialFunctions, and will raise a StackOverflowError if any of them are used (as an operator): JuliaSIMD/LoopVectorization.jl#233. I'm only seeing it in Float16 since that's the first test (which tries a gamma function as an operator).

So this change seems not possible at the moment, unless @turbo is capable of falling back to non-SIMD operations to avoid StackOverflowErrors.

[MilesCranmer]

Raised issue on JuliaSIMD/LoopVectorization.jl#430

[MilesCranmer]

Probably also want to switch to using similar throughout, rather than Array{T,1}(undef, n). This is also needed independent of this change for generality to input array type.

[MilesCranmer]

With JuliaSIMD/LoopVectorization.jl#431 being merged, this should now work. Let's see how stable things are.