Reduce number of allocations

[sethaxen]

Right now there are a lot of unnecessary allocations. Eliminating all such allocations would probably impose too many constraints on the supportable HMC algorithms or require lots of code duplication, but we could still probably gain a lot of performance by at least mutating whenever we know the the vector being overwritten will never be needed for a sample or a sample statistic.

[xukai92]

Mohamed once had such an attempt (#76) but we didn't finally take it because it makes the code harder to read, and at that moment the package was still in fast dev. phase. I guess it's worth a revisit as the package seems to be stable for a while.

[sethaxen]

Thanks for the reference! I think after #228 and #214 are merged, it will be worth revisiting this. In addition to making code more efficient, adopting a mutating pattern should make the code of vectorized NUTS even simpler.

[xukai92]

Just to keep track of things: quick benchmark by @sethaxen (#214 (comment))