Remove Tket dependencies for PennyLane model

[blakewilsonquantinuum]

Currently, to make a pennylane circuit, we construct a tket circuit, extract the ops, and then construct the pennylane circuit from that. This requires tket being a middle man for any circuit conversion. Instead, we can convert to a preliminary diagram compatible with all circuits before converting to the particular backend.

make_circuital function - Built. Needs more testing.
Pulls qubits to the top of the diagram.
Pulls measurements to the bottom of the diagram.
Orders qubits and measurements so they are easy to build a circuit.
Type checks each layer to ensure a valid diagram.
Type checking has been tested on all diagrams from the mc_dataset.
is_circuital - Needs built.
Checks if a diagram is a valid circuital diagram.
Verifies qubits, measurements, and gates are all placed correctly and type checks.
started to build in the draft PR.
from_circuit - Needs built.
Function to build a tket, pennylane, or other circuit using the preliminary circuit.
Refactor tket, pennylane, and torch_quantum to use the above functions for building their circuits.

[blakewilsonquantinuum]

The mypy issue above is unrelated to this PR. It comes from grammar.py when returning NotImplemented and is only flagged in the newest version of mypy (1.14.1 released on Dec 30 2024)

[dimkart]

@neiljdo Another thing we'd like is to make tket an optional dependency, only related to TketModel.

[neiljdo]

@dimkart I've reduced the wall time for model creation (see below):

Before pickle-hack:

After pickle-hack:

There's still some discrepancy between the training times - I'm investigating that next.

[dimkart]

@neiljdo This looks good -- when you finish with the fixes let's attempt another comparison with the previous implementation (sympy symbols + tket in the middle)

[neiljdo]

@dimkart here are the new benchmarking results with the updated PennyLaneModel:

pennylanemodel-w-native-symbol.pdf

The TketModel leaves much to be desired regarding training speed - I can investigate this on a later PR.

[dimkart]

@neiljdo Thanks for the report. The training speed of tket model looks a lot worse compared to previous run, is that normal? Also, how is the performance of the model? Do we get the expected results with the new implementation?

[neiljdo]

@dimkart I've reran the full benchmark and now the TketModel is more comparable to the previous results - I think there might have been some other process that ran during that part of the benchmark.

As for the accuracy metrics, it's pretty much the same for all - it's just the wall times that changed.

[dimkart]

@neiljdo Good work, as long as we move tket to the optional dependencies I think this can be merged.

[dimkart]

Looks good now!