Integration with qiskit-symb

[gcattan]

Issue: #287

This PR aims to integrate pyRiemann-qiskit with qiskit-symb.
There is a performance benefit using qiskit-symb with a small number of qubits (see here)

In practice, in the light benchmark, we can divide the simulation time of QSVC by two using the default parameters.

The solution here was discussed with the library owner in this issue.

There are two elements of complexity in the current implementation:

1. Inside qiskit-symb itself, numpy.einsum:

This PR workaround this by caching the computation (the feature map is always the same in general).

2. In our code (modified from the qiskit-symb issue), in the evaluate method of the new SymbFidelityStatevectorKernel.

[qbarthelemy]

Without any value judgement on the quality of the qiskit-symb code, I am not in favor of adding a package with a single maintainer as a mandatory dependency.

Is it possible to modify the code to add qiskit-symb in pyRiemann-qiskit\doc\requirements.txt but not in pyRiemann-qiskit\requirements.txt?

Functions depending on qiskit-symb would only be called when building the doc.

[gcattan]

Yes, you are right. I am putting qiskit-symb as an optional dependency.

[gcattan]

Thanks for the review @qbarthelemy! Took a bit of work to correct the pipelines.
The light benchmark is still failing as the main branch now also checks NCH with QAOA-CV.
No regression is introduced on the other classifiers otherwise.