Use cache for inference with TextEncoder

[Vincent-Maladiere]

Currently, when we cross-validate the number of components of the PCA used by TextEncoder, or simply when we use TextEncoder in a pipeline being cross-validated, TextEncoder recomputes all embeddings, although they could have been previously computed.

By using a joblib Memory caching, we would be able to compute the embedding of an input X only once in such situations. Note, however, that caching would not work when shuffling the input data.