Move GPU support into Dagger

[jpsamaroo]

I've long been displeased about having to say using DaggerGPU to get GPU support code loaded, which used to be necessary before package extensions existed. I also think that having such important code live outside of the core Dagger repo makes it hard to keep the GPU code up to date and working properly as Dagger itself evolves. This PR corrects these issues by moving all GPU support code into the Dagger repo, and wiring up CI to keep it regularly tested.

Mostly, things are the same as before, with the exception of not needing using DaggerGPU, and with functions being under the Dagger. namespace instead. I've also made a few slight name changes, but I'll keep the original names in DaggerGPU.jl (with deprecation warnings) to ensure existing code keeps working.

Also, Dagger.Kernel no longer automatically synchronizes, and instead users can call Dagger.gpu_synchronize(...) to synchronize Dagger's background queues/streams. Having a forced synchronization basically negated the usefulness of Dagger.Kernel, but was necessary before we did queue/stream tracking. Now you can rely on Dagger to handle synchronization across Dagger tasks automatically, and only need to call Dagger.gpu_synchronize(...) when running GPU code outside of Dagger, to force Dagger to synchronize its background streams/queues.

Todo:

• Add some initial GPU documentation
• (Optional) Add initial OpenCL support code (needs some more testing)
• Prepare DaggerGPU changes (Deprecate DaggerGPU JuliaGPU/DaggerGPU.jl#39)