New Features
- Graph components can now bind inputs from the parent graph. A new node type (ComponentInput) can be inserted into a component graph. This can be dynamically bound by the parent graph, allowing the component to be placed anywhere inside the parent graph. If not bound, then a regular buffer input is bound, which can be queried. See docs for details.
- Exposing nodes for basic component wise math operations (+, -, *, /)

GPU
- Fix occasional lost device when changing input data row count and re-evaluating the graph
- Fix occasional wrong gradients being computed due to mismanagement of GPU buffers

StopGradient Node
- Gradient stop should now apply to entire sub chain instead of just the immediate input. This bug resulted in bad Value calculation in the A2C trainer. That should behave much better now.

Loss Training
- Fix bug causing training buffers to lose training data. Gradient descent training should now work properly with large datasets.

Genetic Trainer
- Reverting adaptive standard deviation raise percent to previous value (1.1 instead of 1.05)
- Removed warning about negative mutation target count when using auto targeting with a graph that contains no conv2d layers.
- Rounding mutation target count when specifying a positive fraction

A2C Trainer
- Fix component loss values always returning 0 with a GPU device
- General stability fixes

PPO Trainer
- Fix component loss values always returning 0 with a GPU device

Agents
- Agents registered with the graph manager will automatically request output they need from the graph.

Graph Manager
- Better determinism when resetting graph manager

Unreal
- BREAKING: Making graph and trainer variables private on the training harness. They can be accessed with GetGraph() and GetTrainer[HTML_REMOVED]()
- Extracting base training harness class from networked training harness to better align with Unity.
- Extending networked training harness with a version specific to genetic training that handles common setup, similar to Unity.
- Training harness only ticks when training is in progress
- Network training harness will save to normal files in addition to appended files during training.

Unity
- Network training harness will save to normal files in addition to appended files during training.