Enable "symbolic compilation" of QNodes

[erinaldiq]

Current behavior

In the current pytket.PytketDevice a compilation pass is applied every time the PennyLane circuit is run (in apply()). This results in a compilation pass every time the variational parameters (or trainable parameters) of a circuit are changed i.e. in a hybrid neural network that is being trained with PyTorch or TensorFlow.

New behavior

When a PennyLane QNode is passed to the pytket.PytketDevice, keep track of what parameters are trainable (for example the ones we request gradients for...) and replace them with fresh_symbol equivalents in the construction of the circuit (e.g. at this line as suggested by @kyasci Kentaro Yamamoto)
Then during the apply() function replace the value of the parameters after compilation.

Questions

• Is this something that can be easily implemented?
• Would it create problems on some backends?
• Who needs this feature?
• Does it make training faster on hardware backends?