[Task] vk: Rewrite our resource binding model

[kd-11]

We don't need to switch descriptors so often if we have some kind of draw state caching. Most of the time we're not really changing anything except a bunch of pointers.
The aim here is to eventually move most work to the GPU including preflight vertex input processing.

• Change descriptor scopes from per-draw to per-state-hash. We calculate hashes anyway for shader live states, might as well make use of that.
• Extend the new descriptor model with EXT_descriptor_buffer to offload processing descriptors to the GPU.
• When DMA passthrough is available, map host mem to the GPU using TTs. Non-trivial since we need to account for boundaries but UTs can help make this watertight.
• When DMA passthrough is available, bypass the vertex input stage using texel buffers by allowing GPU to read data directly from host.
• Offload index processing completely using DGC or DrawIndirect. This approach is so different that it may be a separate VK_1.4+ renderer option. Potential wins are massive here since most users have powerful GPUs but weak CPUs.