* Fix pre-vectorization dataloader iteration. Found with some high
  values of dataset size, memory errors could occur. Using pinned memory
  in dataloader seems to fix this.
* Making use of the file-system sharing strategy for internal
  pre-vectorization, restoring current sharing strategy afterwards.
* Fix cache checksuming to include module and classname of vectorizer
  implementation being utilized.
* Moved internally used dataset for pre-vectorization out as a
  global-level class definition instead of being inside a closure.
* Updated "main" function a bit to be parameterized and spit out some
  useful stats for configuring a training run (number of windows,
  dimension of embedding vector).