improve MC sampling algorithm

[kunyuan]

The following is the roadmap

1. Allow two different types of mesh: a) adaptive mesh from the vegas algorithm, b) user-defined mesh
2. keep track of a histogram for the mesh points.
3. sample the histogram to propose the variable, so that the acceptance ratio is optimized.

Remark:

1. As long as one uses a high-quality mesh, the histogram will be relatively flat. Therefore, one could sample the histogram with a naive MC algorithm (namely, sample the variable uniformly, then reject or accept accordingly to the distribution derived from the histogram).
2. If the histogram is not so flat, the variable sampled from the histogram may be strongly self-correlated. Then the integral evaluation (expansive!) with the variables could be strongly self-correlated, too. To avoid this problem, one could sample the histogram before actual MC integration to learn the autocorrelation time Nt. Then one may choose to do one integral evaluation per Nt histogram samplings (the latter is extremely efficient anyway).