The accumulation matrix for ID metrics should (optionally) be sparse

[tadejsv]

In line with preparation for #4, which uses almost the same kind of matching algorithm as ID metrics, it would make sense to sparsify the matching matrix for ID matrices.

There would need to be some criteria for this - for example, if total number of tracks / number of frames > 10 (i.e. most tracks do not appear on most frames), as only in those cases is this bound to give us any speed/memory improvements. This could also be benchmarked, at the very least on the current integration test tracks.

[tadejsv]

Related to this: sparse matrices seem to be a good option for sparse array construction, as show in #19.

However, if this is the case, I wonder if the best option here would be to simply use numba? Not exactly what it does - but this is still basically just speeding up for loops, so it might be able to help.

[tadejsv]

Done in #23