Gpim

- Fix incorrect enumeration of previous points when using 'dscale' criterion in gpbayes.boptim
- Add utility function for visualizing exploration history in GP-based Bayesian optimization

New functionalities for gp.bayes:
1) Added option to mask certain indices when searching for maximum of the acquisition function
2) Added the additional criterion for selecting the next query point based on the distances to N previous points

Changes:
- Add option to run GP with single or double precision
- Add option to define upper/lower bounds for GP/BO of image data sets
- Add option to run "simulated experiments" (where the full dataset is already available) with GP-based BO

New functionalities:
1. Added GP for vector-valued functions (see gpim.gpreg.vgpr)
2. Created a separate class for GP-based Bayesian optimization (see gpim.gpbayes.boptim)
3. Added expected improvement and probability of improvement acquisition functions (see gpim.gpbayes.acqfunc)
4. Added a batch update option to Bayesian optimization which returns a batch of points separated by distances determined by a kernel lengthscale.
5. Updated verbosity option, which has now three different levels

Breaking changes:
Due to some code refactoring, gpr, skgpr and vgpr are now sitting in gpim.gpreg modules. Hence the import statements like "from gpim import gpr" are not going to work anymore. Instead, use gpim.reconstructor, gpim.skreconstructor, gpim.vreconstructor, gpim.botimizer, gpim.utils.

GPim is a simple python package that provides a systematic and easy way to apply Gaussian processes (GP) to images and hyperspectral data in Pyro and Gpytorch frameworks.