Added support for time step model

[jhogsett]

Support for the Time Step version of the EMA-VFI model has been added to the application. It can be enabled in config.yaml by setting use_time_step to True

Some notes about it:

• It is disabled by default
• I have not had good success using it so far
  • I find I get a different image at the 50% point when comparing both models, even with the original EMA-VFI demo code
  • When deep-interpolating a series, for example, to inflate a video 8X, the final inflated version of the current frame at 0.875 and the next key (real) frame at 1.0 do not line up smoothly, causing "fits".
• Using the 2x (non-time step) model and Deep Binary Split Interpolation, I'm able to get super-smooth inflated videos