Various performance improvements

[KristofferC]

I was playing around with my own version of hyper dual numbers locally and compared against this and saw that there was some performance to be gained here so I thought I should be a good citizen and upstream the various things.

As a benchmark, consider

julia> using HyperDualNumbers, BenchmarkTools

julia> h = HyperDualNumbers.Hyper256(1.0, 2.0, 4.0, 5.0);

julia> f(x) = exp(x) / sqrt(sin(x)^3 + cos(x)^3);

julia> @btime f($h)

On master I get:

julia> @btime f($h)
  167.252 ns (0 allocations: 0 bytes)
3.1313954345926485 + 4.555797620352456ε₁ + 9.111595240704911ε₂ + 3.290579765262758ε₁ε₂

while on this branch I get:

julia> @btime f($h)
  44.958 ns (0 allocations: 0 bytes)
3.1313954345926485 + 4.555797620352456ε₁ + 9.111595240704911ε₂ + 3.2905797652627617ε₁ε₂

[goedman]

Thanks Kristoffer, just merged it.