More unit tests for HMC building blocks

[xukai92]

• metrics
• abstraction related to NUTS
  • tree sampler
  • termination critera
  • full binary tree

[yebai]

More unit tests to resolve the ESS issue (TuringLang/Turing.jl#840).

Some standard unit tests that we should try:

1. compare AHMC and DHMC on a set of hand-coded target distributions (e.g. Numerical/convergence problems with some models Turing.jl#840). This allows us to rule out potential interactions with Bisectors, VarInfo.

2. compare ESS of HMC configurations with mass matrix adaptation, step-size adaptation, u-turn independently, one at a time. This would allow us to rule out interactions between step-size, preconditioning and u-turn termination.

3. compare AHMC and DHMC with vanilla HMC with manual step-size and no preconditioning. This allows us to rule out potential bugs with leapfrog integrator, MH correction etc.