Add layer attributes for all functional models

[mattdangerw]

This proposes a new style for backbone and tasks:

# === Layers ===
self.token_embedding = ...
self.transformer_layers = []
for i in range(num_layers):
    self.transformer_layers.append(...)

# === Functional Model ===
inputs = keras.Input(...)
x = self.token_embedding(inputs)
for layer in self.transformer_layers:
    x = layer(x)
outputs = x
super.__init__(inputs, outputs)

# === Config ===
self.num_layers = num_layers

The main goal is to allow more readable manipulations of the model, e.g. a custom LoRA routine:

backbone = GPT2Backbone(...)
for layer in self.transformer_layers:
    # Use different lora rank for different matrices.
    layer.self_attention.query_projection.enable_lora(16)
    layer.self_attention.key_projection.enable_lora(4)

[mattdangerw]

Draft PR just for feedback for now. I've only touched less than half of the models. Tests will fail.

[mattdangerw]

I also removed unused features from our pipeline model class, just to keep our tasks as clean as possible.

[fchollet]

LGTM, thank you!