This fix now uses a TLSF allocator for GPU memory allocation.
This allows us to make use of proper large blocks, but still
allow small allocations. I think this is a good design; and also
is circumvented when allocating large BOs, so user-space manual
smarter systems will probably not be slower?

The TLSF (ext) implementation is a bit of a hack. We extracted the
relevant source code from Vulkan VMA, fed it to Claude, and tried
to remove all vulkan-related stuff. It seems to work and have
been tested in another repo, but maybe we want to add tests here.

We will see. :^)