Support for float16 training

[ASvyatkovskiy]

This PR enables ability to train FRNN on float16 half-precision floats. Main changes.

1. Introduce a custom MPI_Datatype of 2 contiguous bytes, and save it in the type dictionary:

   mpi_float16 = MPI.BYTE.Create_contiguous(2).Commit()
   MPI._typedict['e'] = mpi_float16

2. Define a custom reduction operation for the new type:

   def sum_f16_cb(buffer_a, buffer_b, t):
   assert t == mpi_float16
   array_a = np.frombuffer(buffer_a, dtype='float16')
   array_b = np.frombuffer(buffer_b, dtype='float16')
   array_b += array_a

3. Register it as an OP:

   mpi_sum_f16 = MPI.Op.Create(sum_f16_cb, commute=True)

4. Introduce a switch in the code based on the Keras floatX variable:

   if K.floatx() == 'float16':
   self.comm.Allreduce(arr,arr_global,op=mpi_sum_f16)
   else:
   self.comm.Allreduce(arr,arr_global,op=MPI.SUM)

Other side changes:

1. Move configuration files to plasma/conf_parser.py, cleanup.