Factor the slab allocator out of the mesh allocator, and make it generic.

[pcwalton]

The mesh allocator has grown fairly sophisticated, with an efficient O(1) allocation strategy, the ability to manage heterogeneous slabs, batched allocation and deallocation, and so on. However, it's currently tied to meshes. In order to scale to millions of mesh instances, I want to allocate bins using the same strategy. So, in preparation for these bin slabs, this commit factors out the generic allocation logic in MeshAllocator into a new generic SlabAllocator type.

The refactored slab allocator is driven by a new trait, SlabItem. Implementing SlabItem requires specifying the size of the elements that are to be stored in the slabs, as well as the type of the key used to retrieve those items. A single slab allocator can be directed to manage heterogeneous slabs by implementing the associated SlabItem::Layout type.

This PR also refactors slab allocation and deallocation to be explicitly transactional. This means that users of the allocator no longer need to have special logic to deallocate empty slabs and reallocate slabs that changed in size. To perform batched allocation, code that uses the allocator calls stage_allocation to retrieve an AllocationStage object, calls AllocationStage::allocate repeatedly to perform all necessary allocations, and finally calls commit; slab reallocation happens efficiently under the hood. Likewise, to perform batched deallocation, the client code calls stage_deallocation to create a DeallocationStage object, calls DeallocationStage::free to free all necessary objects, and then calls commit to free slabs that are have become empty.

Along with the forthcoming PR to add SparseBufferVec, this PR is a prerequisite to bin slabs. Moreover, it makes the mesh allocation code considerably cleaner and easier to understand.