Return correct output shape for R(2+1)D

[jsjason]

Our implementation of the R(2+1)D model outputs tensors of different shapes, depending on the start_index and end_index model parameters given at class instantiation. However, the output_shape() method of R2P1DRunner is set to always return (10, 400), which is the output of the final layer in R(2+1)D. In fact, this is an artifact from #58, in which we made output_shape() a static method; the method cannot access internal attributes like self.start_index and self.end_index to return the correct output shape. We need to change the implementation so that output_shape() can return different values according to model parameters.

One possible way to do this would be to retrieve model parameters from the input pipeline configuration and pass the parameters to output_shape(), similar to what we are doing in benchmark.py:

for step_idx, step in enumerate(pipeline):
  ...
  process_runner = Process(target=runner, args=...,
                           kwargs=step) # <-- right here

This way, we can let the output_shape() method determine which shape to return based on model parameters, while still keeping output_shape() as a static method.