Map Strings to Tensorboard Projector

When projecting the embeddings from the TFRS Multitask Tutorial, the StringLookup layer transforms the movies and users into integers. These integers are projected. Users are fine this way, but movies would be more meaningful of a projection if you could rapidly see the string that was mapped.

With that being said, it's relatively unclear how to map strings back to play nicely Tensorboard Projector. Many of the embedding commands no longer work for TF 2.x, and that model is wrapped in Sequential such that its elements cannot be simply extracted.

I would recommend having the ability to upload the dictionary either as a Python dictionary or as a 2 column .csv/.tsv with the integer left and the string right. Command line would be fine but UI could also have a "map strings from .csv" function as well.

Start by reproducing the TFRS Multitask Tutorial projection in TensorBoard Projector and inspect the current embedding workflow; the issue notes that TensorFlow 2.x commands and Sequential extraction are problematic. Done means a supported way to load an integer-to-string dictionary through Python, CSV/TSV, CLI, or UI and display movie strings in the projection.