Get the examples running again.
The compute example has been simplified to use only one buffer.
A new check has been added to ensure the `max_bind_groups`
device limit is properly set.
This PR removes optional backend features in favor of run-time
selection at the `request_adapter` entry point. Adapters from all
supported gpu backends on the platform are reported.
It also makes gfx-rs dependency to be public (for now).
