Context is flutter/engine#55694

So I think this is a problem in the HAL design. We expect to be able to have a pipeline synchronously for encoding but

1. we can only get the pipeline on a particular thread for GLES
2. we don't actually need the pipeline until we render - which always happens on the react thread.

I think it would be something to punt to a new PR, but I would consider change the HAL design such that SetPipeline accepted the pipleine future. The Vulkan/Metal implementation would call WaitAndGet immediately while the GLES implementation would defer that until encoding.