fix #1362. prefetch HDF5DataLayer

[pclove1]

@jeffdonahue , @shelhamer

This PR attempts to fix #1362.
I highly recommend you to read #1362's description first to see what this PR tries to achieve.
To be short, we want to prefetch HDF5 files and avoid excessive memory usage by reading HDF5 files partially.

The below is the summary of modifications made on top of #1362.

• override DataLayerSetUp() instead of LayerSetUp() following the existing design
• Blobs are N-D arrays now since Blobs are N-D arrays (for N not necessarily equals 4) #1970
• wait until a thread is joined in when destructing HDF5DataLayer
• preserve the ability of shuffling files
• add a unit test case where a batch is loaded from interleaving HDF5 files to cover the found error on the index math.
• plus a few minor bug fixes

Caveats

• don't shuffle rows in a HDF5 file due to the introduction of partial reading
  • I believe this is also true for LMDB/LevelDB cases where the users are supposed to shuffle the data when they construct them.
• If someone needs to access the same HDF5 file concurrently, they need to make sure to have a thread-safe version of hdf5 library. (See this)

[ronghanghu]

Rebased and cleaned up in #2892. Thanks the original authors @jeffdonahue and @pclove1. Authorship preserved in commit messages.