Implement dataset and dalaoader classes

[gutzcha]

Implement dataset class for training the MAE model.
Model input:
N x T x F x D
where:

• N = batch size (?)
• T = number of frames per window (64 representing ~2 seconds)
• F = number of features/landmarks (478 using mediapipe's face_mesh)
• D = feature dimension (2 x, y)

The dataset was saved in json files with the following fields:

• id: frame number
• results: pose data
  • list of tuples with length=478 landmarks, each tuple contains x and y coordinates
• timestamps

see json file sample:
extracted pose sample