Non-transposed dense oracular extractors now directly fill the cache.

This uses a shared memory pool where the existing slabs are defragmented
to make a contiguous allocation of available cache memory. This is used
in a single call to the HDF5 library with hyperslab unions for the new
slabs to be populated. The aim is to reduce the overhead from separate
calls to the HDF5 library, a la the primary sparse extractors.

Note that this optimization is only available when the target dimension
corresponds to HDF5 rows. If transposition is required, we do separate
calls for each slab, we can't afford to hold all slabs in a separate
buffer before performing the transposition into the cache pool.

As a result, the oracular DenseMatrix extractors are split into two
subclasses depending on whether the extraction requires transposition.
We do the same to all other DenseMatrix extractors for consistency.