Remove fragmentation of slab memory due to avg. tuple size growth

In a scenario, when average size of a slab item gradually changes, our slab allocator fragmentation ratio grows, and eventually a lot of slab classes are unsuitable for use.
Today this can only be solved by a server restart.

We should:
- investigate options and maybe replace our slab allocator altogether, e.g. with a variant of Doug Lea allocator, jemalloc, etc.
- otherwise, find a way to reuse/defragment slab classes.