Let actors use GPUs.

[robertnishihara]

This is a quick hack to allow actors to use GPUs so we can figure out what the right API here is.

In this PR:

• When creating an actor, the decorator can take a number of GPUs (and CPUs, although that field is currently not used for anything), e.g., via ray.actor(num_gpus=2).
• Actor methods can call ray.get_gpu_ids() to get a list of the GPU IDs that the actor is allowed to use. The idea is that this could be used in the constructor of an actor (which defines a neural net) via something like

os.environ["CUDA_VISIBLE_DEVICES"] = ",".join([str(i) for i in ray.get_gpu_ids()])

• GPU management here is done entirely outside of the local scheduler. The worker or driver that creates the actor will interact with Redis to see how many available GPUs each local scheduler has, will take some of them for the newly created actor, and will mark them as used. If there aren't enough available GPUs, it will raise an exception. TODO: this needs to be unified with how resource management is done in the rest of the system.
• GPUs are never freed once an actor consumes them. More generally, actors are never garbage collected. TODO: this needs to change.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/56/
Test PASSed.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/58/
Test PASSed.

[pcmoritz]

This is duplicated

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/60/
Test PASSed.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/61/
Test PASSed.

[AmplabJenkins]

Merged build finished. Test PASSed.

[AmplabJenkins]

Test PASSed.
Refer to this link for build results (access rights to CI server needed):
https://amplab.cs.berkeley.edu/jenkins//job/Ray-PRB/62/
Test PASSed.