Model Factory Everything is a block

[lucidtronix]

What
New model factory interface everything is a block.
ResNet, densenet, attention, weight sharing, variational bottlenecks, normalization and regularization are all blocks.

U_nets handled with denseDecoder style blocks or keep the automatic reflection?

How to specify modality specific blocks and aggregated blocks?

Why
Hard wired architecture details
We want to handle variable sized input, language inputs
We want to handle arbitrary internal connections, as we need for many attention based models, not just UNet style skips.

How
--blocks argument takes list of strings from the BlockType Enum.
Model factory now just iterates over blocks.
implement in parallel to existing model factory.

Acceptance Criteria
Default model parameters yield models with performance similar to master.
Attention, weight sharing and pooling concept level blocks allow dynamic shaped inputs and outputs.

[lucidtronix]

This also should address #190, serializable encoders and decoders, streamline using distance with hidden inference as a loss.

[lucidtronix]

And #385, Transformers and attention as model blocks

[lucidtronix]

First pass is in #393,

• Model factory can return encoders & decoders as models,
• Encoders can be model
• External Model files can be model blocks
• Color nested models in architecture diagram
• Tokenizers are encoders