Dota2 benchmark regression

[kvark]

Somewhere between https://gfx-rs.github.io/2018/08/10/dota2-macos-performance.html and today we regressed the benchmark significantly. We used to get 33.9 / 3.5 (in the table), but now we are closer to just 30 (in immediate mode).
Here are the things to try:

• new swapchain model is less efficient at fullscreen? Tried with "-sw" parameter, no change.
• Dota doesn't discover some of the features we have, it needs to support VK_KHR_portability_subset
• on macOS 11 we aren't using some of the Metal features we are supposed to use
• we are locking too much, and that hurts on macOS11 beta, possibly because of parking_lot regression? Tried parking_lot-0.6, no change.
• macOS 11 beta has some strange scheduling issues. Maybe background apps interfere too much? 🥇 disabling Firefox in the background gets us from 30 to 32fps in average.
• setting labels on internal encoders is slow - hid them behind a flag now, not sure if changed any metrics
• ExactSizeIterator bounds are confusing the compiler and making us heap-allocate more?
• Rust optimization regression?
• creating and filling MTLRenderPassDescriptor is too slow - implemented a pool for reusing them in gfx-rs/gfx@bbf55ff
• binding descriptor sets is still to slow - rewritten in gfx-rs/gfx@15e456b
• too many redundant depth-stencil changes - fixed in gfx-rs/gfx@0d581a6
• redundant scissor changes - fixed in gfx-rs/gfx@f0051da
• we are doing many redundant passes - investigated, no clear picture of why that would happen
• we are rebinding too many resources - testing a fix but it gave no performance difference. We are binding them in ranges, anyway, so Metal run-time possibly does this for us.
• we are using mutable pixel views too often - fixed in gfx-rs/gfx@22e2f4d

Related to gfx-rs/gfx#3378, gfx-rs/gfx#3382, gfx-rs/gfx#3381, gfx-rs/gfx#3383

[kvark]

Dota doesn't discover some of the features we have

I just noticed that it does indeed print the following lines when ran on gfx-portability:

Vulkan physical device (0): does not support Metal depthSampleCompare.
Vulkan physical device (0): using transform constant buffer: false
Vulkan physical device (0): supports shader clip distance: true
Vulkan physical device (0): using secondary command buffers: false

I imagine working around the comparison samplers has a solid GPU performance cost, and on Intel GPU it would show up as an overall regression. So I believe this issue is blocked on ValveSoftware/Dota-2-Vulkan#351