Optimized the creation of CommandEncoders on WebGPU

[mvaligursky]

The rendering on WebGPU used to create a new command encoder for each render pass, and also for other operations between the passes, like buffer upload / download, query resolves and similar.

Now this has been minimized to 2-3 a frame, depending on how the scheduling of dynamic buffer uploads needs to execute.

The hope is that this will minimize the gaps on the gpu between render passes, which this likely contributes to, at least on Apple HW. Follows this advice from Apple engineer:

Each command encoder creates a MTLCommandBuffer and MTLCommandBuffer requires synchronization
back to the unified memory on Apple Silicon and iOS. So this is much much slower in general than using
a single MTLCommandBuffer (1 GPUCommandEncoder) and creating several render or compute passes.

This is the before capture, where the gaps can be observed. It seems the hea dof mini-browser is broken at the moment, and so I cannot capture after screenshot currently :(

Note especially the gaps between those tiny bloom blur passes, where the gaps are significantly larger than the passes themselves.