Add cached evaluation of tensor networks

[DNA386]

Add caching option by default for models that evaluate tensor networks, as per #185
Also bonus include the Sim9 and Sim9Cx ansatze from the inTask repo.

A non-cached option is also included in case the circuits are small enough that caching introduces unnecessary overhead, but must be actively selected by the user. By default this will reproduce the previous behaviour.
The user can select this as a kwarg when initialising the model. Note: the pathfinder must also be provided as a kwarg if resuming from a checkpoint - the checkpoint currently will contain the saved paths, but not the information required to re-build the same fallback path finding algorithm.

By default, the paths will only be cached in memory. The user can chose to either save them to the checkpoint, or to a separate file by supplying a configured CachedTnPathOptimizer instance.

[neiljdo]

Hi @DNA386 thank you for your work! 🙌🏼 I just came back so I can only do a surface-level review at the moment - kindly take a look at my initial notes. I'll have a closer look soon.

One thing we would like to have are tests specific to the classes/functions implemented in the `lambeq/training/saved_tn_optimizer.py'.

[dimkart]

We can make this configurable (i.e. enable/disable storing the paths in the checkpoints somehow).

This initial proposal currently saves cached paths to the model checkpoint, but this is not desirable if the datasets are very large as this will result in a lot of duplicated information (assuming checkpoints are saved per epoch). The paths can typically also be reused when training a different iteration of a model on the same dataset (with the same anstaz), which this system doesn't capture. A more natural way to implement is to have an independent checkpointing system for the paths, that overwrite the previous checkpoint each time a new path is added, however this means the user needs to specify and track the checkpoint filepath separately from the model. In this case, it would not be possible to have caching by default; the user would need to explicitly opt-in by providing a configured TnOptimizer.

[dimkart]

@neiljdo Let's run our benchmarks on this PR and compare speeds with current branch.