[ENH] Have a subset of sampler enabling sampling in large dataset

[glemaitre]

We got a couple of issues, notably with SMOTENC, where large datasets drive to a MemoryError.

Here I will add a couple of points that could be addressed in the future:

• Check in the class SMOTENC if converting a dataset from sparse to dense is required (SMOTENC MemoryError #752, SMOTE-NC sampling_strategy='not majority' MemoryError: Unable to allocate 135. GiB #768, MemoryError SMOTENC #688, Benchmark for dataset size before Memory Errors on SMOTENC resampled dataset creation #667)
• A subset of the sampler could be implemented in Dask. We should probably prototype in imblearn before to contribute it upstream (Samplers / pipelines for imbalanced datasets dask/dask-ml#317)

[arisingh8]

Was there any progress on this issue? It remains a breaking problem on large datasets for the current release.