Support sparse data formats

[mih]

At the moment a constant sampling rate is assumed, due to our choice of input format (one sample per line, no time column). However, sparse data formats could be supported by converting then to a dense timeseries array on the fly.

Alternatively, we could modify the algorithm itself to operate on sparse timeseries (with potential gains in terms of memory load). That should not be a major change, but I guess the required effort is larger than the convert-on-load solution.