V2 Encoder Architecture

[jakehlee]

This is a draft PR for an updated SpecTfEncoder architecture. I am calling it V2 to avoid breaking backwards compatibility.

• The banddef, or the wavelength grid of the spectra used for the learned positional embedding, was previously a static parameter of the model architecture. This is because we expected this value to not change at the instrument level (e.g. for EMIT). However, we now know this is a parameter that does change over time, especially for other instruments. In the V2 architecture, the banddef is now an input feature, not a model parameter. This means that a different banddef can be passed per-spectrum, or a single banddef can be passed for the entire batch.
• A new optional mask argument is added to support variable-length input sequences in the future. This would allow spectra with entirely different spectral grids (e.g. AVIRIS and EMIT) to be fed within the same batch for specific inter-instrument training.
• The "flat" aggregation method, which also locks the sequence length to a fixed value, has to be removed. Without prior knowledge of the sequence length, "flat" aggregation cannot guarantee a static input feature dimension for the classification layer. Further work/research has shown that "max" and "average" aggregation does fine, with some preliminary evidence showing that one of the tokens starts acting as an aggregation token.

Topics for discussion/comment:

• It could be argued that, instead of branching to a V2 architecture, we should update the default SpecTfEncoder implementation and port over the old model weights to the new architecture. This is a bigger lift but it's technically possible with some work.

I am currently using this implementation on a research task, if that task goes smoothly and we resolve the point above we can open the PR to be merged.