Account for multiple nodes in optimizer cost

[gmittal]

Fixes #28.

Tested

sky cpunode (expect 1 node AWS m4.2xlarge):

I 02-07 21:07:17 optimizer.py:211] Defaulting estimated time to 1 hr. Call Task.set_time_estimator() to override.
I 02-07 21:07:17 optimizer.py:316] Optimizer - plan minimizing cost (~$0.4):
I 02-07 21:07:17 optimizer.py:330]
I 02-07 21:07:17 optimizer.py:330] TASK     BEST_RESOURCE
I 02-07 21:07:17 optimizer.py:330] cpunode  AWS(m4.2xlarge)
I 02-07 21:07:17 optimizer.py:330]
I 02-07 21:07:17 optimizer.py:284] Considered resources -> cost
I 02-07 21:07:17 optimizer.py:285] {AWS(m4.2xlarge): 0.4, GCP(n1-highmem-8): 0.473212}
I 02-07 21:07:17 optimizer.py:285]

python examples/multi_hostname.py (expect 2 node AWS m4.2xlarge)

I 02-10 17:15:20 optimizer.py:211] Defaulting estimated time to 1 hr. Call Task.set_time_estimator() to override.
I 02-10 17:15:20 optimizer.py:317] Optimizer - plan minimizing cost (~$0.8):
I 02-10 17:15:20 optimizer.py:332]
I 02-10 17:15:20 optimizer.py:332] TASK                      BEST_RESOURCE
I 02-10 17:15:20 optimizer.py:332] Task(run=<fn>)            2x AWS(m4.2xlarge)
I 02-10 17:15:20 optimizer.py:332]   nodes: 2
I 02-10 17:15:20 optimizer.py:332]   resources: {AWS(None)}
I 02-10 17:15:20 optimizer.py:332]

[gmittal]

Question: should the egress cost also be multiplied by num nodes?

I assume we usually transfer to the head node and then copy from head node -> workers rather than from previous node in the DAG to all the worker nodes.

EDIT: Added FIXME for now.

[concretevitamin]

From a quick look, BEST_RESOURCE should show <n>x <resource str>.