Feature Idea: Support for Keras Functional API

[davidrsch]

Is your feature request related to a problem?

The current implementation is built on keras_model_sequential(), which limits users to a linear stack of layers. This prevents the creation of modern, complex architectures with non-linear topologies like residual connections (ResNets), multiple inputs/outputs, or shared layers (Inception-like models).

Describe the solution you'd like

A new model specification factory, create_keras_functional_spec(), that allows users to define a model as a directed acyclic graph.

Instead of a simple ordered list, the layer_blocks argument would allow each block to specify which preceding block(s) it takes as input via an input_from argument.

Hypothetical Syntax:

create_keras_functional_spec(
  model_name = "resnet_like",
  layer_blocks = list(
    input = input_block(),
    path_a = dense_block(input_from = "input"),
    path_b = conv_block(input_from = "input"),
    concat = layer_concatenate(input_from = c("path_a", "path_b")),
    output = output_block(input_from = "concat")
  ),
  mode = "regression"
)

Acceptance Criteria

• A new factory function create_keras_functional_spec() is created.
• The internal fitting engine (generic_keras_fit_impl or a new variant) is updated to parse the input_from dependencies and construct the Keras functional model graph correctly.
• The new factory supports architectural tuning (e.g., num_{block_name}).
• The function is documented with a clear example of a multi-path model.
• Unit tests are added to verify that a simple functional model can be created, fit, and tuned.