Disk allocation should use multiple region sets

[jmpesp]

test_disk_backed_by_multiple_region_sets creates six 10 GiB crucible datasets, then tries to create a 20 GiB disk. This test currently fails because the the current allocation algorithm does not split a disk across multiple sets of datasets - it queries for the three crucible datasets with the most available space and fails if there isn't room in them for the virtual disk.

Change the allocation algorithm to use as many datasets as is required to fulfill the total disk size.