Force DenseWithSparseWeights to have at least one entry per row

[JEM-Mosig]

Description of Problem:

Whenever DenseWithSparseWeights has completely zero weight rows (which can happen if sparsity is high), it effectively reduces the layer size, which is pointless. So we should prevent this from happening.

Overview of the Solution:

Examples (if relevant):

Blockers (if relevant):

Definition of Done:

[JEM-Mosig]

It might be even better to use LocallyConnected1D layers. Treating this as the same issue.

[JEM-Mosig]

LocallyConnectedDense layers are slow and buggy (implementation!=1 doesn't work) in current Tensorflow version, so I backtracked on those changes. We're now using our own RandomlyConnectedDense layer.

I'm also forcing every input to be connected to at least one output, because it doesn't make sense to ignore inputs randomly.

[JEM-Mosig]

changes to the sparse layers definitely stabilize TED's performance at low densities:

Solid lines show performance at fixed density (see legend) with my modified sparse layers. Dash-dotted lines show the equivalent on the main branch.

The solid green curve is the fully dense layers one. The orange curves are the 20% density (or 80% sparsity) which are our default - for those my changes don't do anything.

The dash-dotted red curve is further to the left, because on main sparse layers are allowed to drop inputs or outputs and thus can have fewer trainable weights.