render/vulkan: Recycle memory maps for stage spans
This is an equivalent to !4718 (closed), but recycling memory maps of stage spans instead.
Compared to the other MR on my machine in with an evil shm client (oversampling weston-simple-shm), this trades a slightly lower CPU utilization (25%->20%, likely just due to the staging buffer having tighter packing) for higher GPU load (6-8% -> 18-20% according to amdgpu_top).
Comparisons from someone with a dGPU would be nice.
Edited by Kenny Levinsen