improve performance of read_object with idx by reading in whole objec…

[lvarriano]

…t, then slicing

Significantly improves performance of read_object when using the idx parameter (see #29) at the cost of slightly increased memory use (up to about 150 MB for one channel of a raw cal file of 15 GB, but typically/always a single channel is loaded at a time). Depending on specifics of file, such as number of events, number of idx, and length of gaps between idx, this can improve performance by up to ~100x.

[jasondet]

• add a flag like use_h5idx to read the original way in case the user cares more about limiting memory than speed
• add blurbs in the documentation about this issue, linking to the issues where this is discussed, and comment on the fact that the "fast" version can't be used with obj_buf

[lvarriano]

This works for a single file without a problem, but once the unit tests try to read in multiple files, some issues with array views and indexing crop up. This turns out to be significantly more complicated to do because of how things are structured when reading in multiple files... I will close this pull-request for now until a better implementation can be worked out.