[ENH] Make minirocket capable of taking unequal length collections

[TonyBagnall]

part of #1699 makes MiniRocket capable of unequal length and deprecates the MiniRocketMultivariateVariable class. This will be rolled out to the other convolution based transformers, also giving associated estimators capability:unequal_length: True tag.

The main issue is you cannot pass a both 3D numpy (equal length) and list of numpy arrays (np-list for unequal) to same numba parameter described by decorator. There are two locations that use numba functions that have to be changed:

1. _fit_biases: this uses series length internally here

            _X = X[np.random.randint(n_cases)][channels_this_combination]
            A = -_X  # A = alpha * X = -X
            G = _X + _X + _X  # G = gamma * X = 3X
            C_alpha = np.zeros(
                (n_channels_this_combination, n_timepoints), dtype=np.float32
            )

so my solution is to split it into two functions _fit_biases_numpy and _fit_biases_list. Currently the second is not numba, since I dont think you can easily pass a list of numpy (could very well be wrong). It is not computationally intensive 2. static _transform this loops through each instance transforming it. My solution is to take this loop out of numba and have a new function _single_case_transform where we pass the case, etc

    _X,
    features,
    n_channels,
    n_timepoints,
    n_dilations,
    n_features_per_dilation,
    dilations,
    n_channels_per_combination,
    channel_indices,
    biases,
    n_kernels,
    indices,

an alternative would be to just remove the decorator typing (not sure if that works) or just have two separate private functions. I'll benchmark times, but atm it looks like it slows things down too much, I'll post graphs below