memory allocation to avoid mem leaks

* added an IndexIVFScalarQuantizer type that offers an
  intermediate compression between IVFFlat and IVFPQ

* support removal of indices in IndexIDMap / IndexFlat combination

* various fixes in GPU code

avoid reading beyond the end of an array in fvec_L2sqr and related functions

- better selection of training sets for PQ and preprocessing
- GPU parameter object
- IndexIDMap fixed
- fixed redo bug in clustering