Convolution sampler and rounder

[blowfish880]

This adds convolution techniques to samplers and randomized rounders, i.e. combining discrete Gaussian distributions to obtain different ones. For samplers, this is useful to reduce memory overhead and setup costs for wide distributions (large sigma). For rounders, this is useful to reducing the task of randomized rounding to the sampling problem, which can be faster than rejection sampling or Karney's algorithm (depending on parameters, hardware, etc).

[malb]

Can you rebase on current master?