Generalize resource management

[w568w]

Currently, node resources are simply understood as NVIDIA GPU resources. Any scheduling revolves around one to multiple GPUs. All resources are abstracted as a Device struct that holds information specific to NVIDIA GPUs. It is not possible to specify other resources or submit a task that requires zero graphics card.

Device needs to be further generalized into something like Resource, which should at least include memory resource for each node.

p.s. as for CPU, I believe the implementation is easier than that of memory; moreover, once memory resource management is implemented, it is easy to reuse the logic for CPU, so we can temporarily set it aside for now.