Blender Git Loki

Half Floats: Fix CUDA rendering

The issue was not in the kernel, but in the bindless mapping instead:
it was allocating space for 4096 slots only, causing writes past the
array boundaries to begin with and making kernel unable to fetch proper
bindless texture object to end with.

Using allocations based on the slot index plus some padding to avoid
too much re-allocations now. A bit weak and maybe allocating all slots
at once is the better alternative, but should let us to work fine for
the time being.