Use memory-mapping with HDF5 files whenever possible

[astrofrog]

This seems to work. I tested it out with @maartenbreddels's sdss_dereddened.hdf5 file and the scatter density plots worked ok. They were slightly sluggish but not dramatically.

@maartenbreddels - when you get a chance, can you review this? Thanks for the example implementation!

Open questions:

• Should we only do memory mapping for arrays over a certain size, performance-wise?

• Does this benefit from any disk I/O caching or does every array access result in an I/O operation?

I also noticed that the auto-merging isn't quite right at the moment, so I'll try and work on that soon, maybe in a separate PR.

[maartenbreddels]

Yes, I'll take a look next week!

[maartenbreddels]

LGTM.
I wouldn't worry about the array size, for small files the performance isn't an issue anyway. The only issue is actually OSX, their mmap support is really bad, as is their diskcaching strategy. Maybe we should have an config option, or keyword argument for when you don't want mmapping (for benchmarking, or when an issue arrises)

[astrofrog]

@maartenbreddels - just to check, what kind of issues do you see on MacOS X typically? Just performance, or actual crashes?

[maartenbreddels]

Performance only.