Reduce memory consumption of very high-resolution merges

[nvictus]

This PR addresses a high memory consumption issue when a large number of very high resolution coolers are merged. It should improve the performance not only of cooler merge but also of cooler cload pairs and cooler load.

In pre-calculating offsets to use for the merge execution plan, we were loading (and concatenating) all bin1_offset indexes into memory. This isn't an issue for typical coolers, but can become prohibitively large for many inputs at high resolutions, where a single index vector can be ~2GB in size at human 10bp resolution.

• Now we use lazy HDF5 datasets and load each bin1_offset index incrementally during merge execution planning. This results in a drastic improvement for merges involving e.g. 100s of datasets.
• We also expose the merge buffer argument to cooler cload pairs and cooler load, and the max-merge option to cooler load, to give the user more flexibility in controlling maximum memory consumption during the actual merge epochs.