Repeatedly creating and freeing a large allocation can fail on a 32bit system if not enough address reuse occurs

[RalfJung]

This now fails on a 32bit Windows targets, but I think it's just a matter of randomness that we haven't seen this fail before:

    for _ in 0..10 {
        drop(Vec::<u8>::with_capacity(512 * 1024 * 1024));
    }

The trouble is that there is some chance we'll want to assign new unique addresses to each allocation, and of course that won't work if there's only 4 GB of space available in total.

We should probably either always try to reuse "sufficiently large" allocations, or always try reuse if we are otherwise running out of address space.