Support Cycling a CombinedStreamingDataset

## 🚀 Feature

The ability to provide length as an input argument to the CombinedStreamingDataset such that the epoch length is dissociated from the number of samples in the dataset. Same as ParallelStreamingDataset.

### Motivation

I want to create a CombinedStreamingDataset that is the weighted combination of StreamingDatasets but be able to specify the number of training steps/cycle the CombinedStreamingDataset arbitrarily. As discussed with @tchaton.

Related to #524

### Alternatives

Not sure if this would work but conceptually one workaround might be to wrap the CombinedStreamingDataset with the ParallelStreamingDataset? e.g.
```
ds1 = StreamingDataset(...)
ds2 = StreamingDataset(...)
cds = CombinedStreamingDataset([ds1, ds2], weights)
pds = ParallelStreamingDataset([cds], length=100)
```

Start by reading the CombinedStreamingDataset and ParallelStreamingDataset implementations, then review the context in related issue #524. Determine how ParallelStreamingDataset accepts length and how cycling should affect weighted combinations; done means CombinedStreamingDataset can accept an arbitrary length and cycle for the requested training steps.

Written by the indexing model from the issue text.