sharded_slab: Support alignment

[avdgrinten]

Alignment works like this:

• If an object's size fits into a slab after padding it to alignment, allocate it from a slab. Slabs are naturally aligned, so this is correct.
• Otherwise, if the alignment is smaller than the chunk alignment, allocate it as a large object (i.e., allocate a single object chunk that has the chunk header at chunk_boundary and the object at chunk_boundary + alignment).
• Otherwise, the object is called overaligned. In this case, allocate the object itself at a chunk_boundary and store the header out of band. Overaligned objects are the only objects that have addresses that are multiples of chunk_boundary, so they can be detected in the deallocate() path.

[avdgrinten]

Actually, for non-power-of-2 buckets, this does not work. Even if the allocation size if a multiple of alignment, the bucket size does not need to be. So this needs some re-thinking.

[avdgrinten]

Note to myself: copy/paste mistake.