[MNT, DOC] Accelerating deep testing

[hadifawaz1999]

Fix #1761

• Changes test params of deep classifiers/regressors/clusterers so that testing is faster
• In deep specific tests, use the test parameters, as they were not used before, to accelerate testing
• Fix random state deep clustering test, it was looping infinitely before because there was a bug, it was considering the encoder/decoder as layers before, not its fixed and ensures encoders and decoders are testing separately
• Add possibility to use dummy clusterer in deep clustering, for the clustering in latent space, use it as test parameter to accelerate testing especially in no numba label. Another reason for this addition is that with test parameters of AE based clusterers in deep learning, if not trained long enough, kmeans may not converge and CI will fail with "ValueError: Unable to find a valid cluster configuration with parameters specified (empty clusters kept forming). Try lowering your n_clusters or raising n_init." so it will be a pain to find the best seed, so replacing with dummy clusterer is best option
• Fix Encoder Network issue with Lambda layer, using now internal attention layer, as the original authors implemented the network before keras introduced Attention layers. The math isnt 100% the same but the idea of attention is the same
• Doc issue with AEResNet