Changes to support packaging

[kurman]

1. Load configurator.py relative to training script
   ===================================================
   Right now configurator.py is loaded from CWD, however training script can be invoked as a python module. We can either convert configurator to a module or load it relative to train.py script. The latter is simpler and works.

2. Allow explicit dataset path arguments
   ========================================
   Similar to previous issue, dataset files may exist in other location relative to CWD so we can add path parameters that can be overridden using existing configuration mechanism

[karpathy]

This commit does two things: the thing you mentioned but also it introduces new variables for train/val paths...