Variable size serving input

For some reason mlflow serve does not accept the variable sized input of the BruteForce layer.

Things tested so far:

• Setting input spec
• Calling predict before log_model with various different sizes
• Wrapping BruteForce by input layer (functional API) with specific input_shape. See https://gitlab.sauerburger.com/frank/movie-recommender/-/blob/80b23288e95f3fa65a1122db54ef1164a46547f5/movies.py#L327

The situation might change when switching from BruteForce to ScaNN layers.