Allow partial processing of read datasets

[johnlees]

For large read datasets, their size (especially plus a large countmin table) may exceed device memory. However they can happily be loaded (asynchronously) in blocks, as each read is independent.

This change will require:

1. Remove the read interleaving, and loading into shared. A sync block load can be used to load all the reads for a warp into shared directly and more simply.
2. Pin the host memory in the DeviceReads class
3. Add a loadBlock method to DeviceReads which loads up to a specified size using memcpy async on a non-default stream
4. Launch the kernel on a second non-default stream.
5. Iterate through loadBlock, waiting on kernel completion.

The block size will be (device memory - size of countmin table) / 2 - epsilon

[johnlees]

Iterating over k-mers is a bit irritating when this is used. If the buffer loading is only to be done once, then signs for all k-mers need to be maintained on the device (fine) as well as the countmin table (not fine). To start with, let's just try keeping the iteration as-is, and loading the buffer on and off for every k-mer length and see how that copes.