[meta] GLES3 backend restrictions

[kvark]

This is spawned from #1510. We are trying to express all the additional restrictions in wgt::DownlevelCapabilities, but some things we don't know how to express, yet. So this issue is meant to track them.

• GL is inherently single-threaded. It's unclear if it's valid to access an EGL context from multiple threads even if they are externally synchronized. It seems like it could work on some drivers but not the others.
• each texture is only allowed to be used with at most one sampler (combined) in all the shaders making a pipeline.
• unable to select mipmap levels and array layers for the views of sampled textures
  • we can modify the texture state to limit the mipmap levels, but this would affect all the views into this texture at once
  • the array layer restrictions could possibly be implemented on the shader side, but this is work
• unable to interpret 2D array texture as cubes - Handling cube textures in GLES backend #1614
  • currently we detect if it can be a cube map, and make it a cube map. This is a hack we deserve, but not the one we need.
• base instance is not supported
  • API side is fixed by re-binding the instance-rate buffers with different offsets
  • shader side needs Base instance semantics on GLES backend #1573
• base vertex is not supported
  • exposed as wgt::DownlevelFlags::BASE_VERTEX
  • perhaps, we can still work around it in a way similar to base instance?