[meta] Performance opportunities on the horison

[kvark]

Merely porting the existing WR device to gfx-hal is just a start. The real fruits will come once we make deeper changes to the WR architecture to take the benefit of modern graphics APIs. This meta-issue is tracking those architectural changes:

• Parallel recording - Record command buffers ahead of time #188
• Specialization - Use specialization constants for WR features #211 (Most of the features are behind a spec const now expect dual source blending which requires to many shader changes)
• In-place instance data - Write out instance data in-place when forming batches #219
• Share descriptor sets between instances and frames - Share descriptor sets between instances and frames #236 (This is mostly done, which we miss from the issue is upstreaming the concept to WR)
• Pipeline cache - Use gfx-rs' pipeline cache object #241 (Vulkan backend already uses this. Metal backend can't write/read a paritcular cache, but internally can manage it between resizes. DX12 backend lacks this feature. )
• Better memory management - Better memory management #255
• Array render passes - Match passes in WR with render passes in the native API #269
• Uber-pass frame rendering - Uber-pass frame rendering #273
• Persistently mapped GPU cache - Persistently map GPU cache data #271
• Manage descriptor sets on the render backend - Manage descriptor sets on the RenderBackend side #334
• Remove all uniforms - Remove all the uniforms #345
• Dynamically allocate render targets - Dynamically allocate render targets #348