run haproxy on multiple nodes

[laszewsk]

only done if single node works

node = computer with n gpus

for each node 
   for each gpu
      start one tfs for the gpu

for each node
   start a haproxy

now figure out what the benchamrk looks like. e.g. we likely run it on each node, but then ther eis no advantage of running multiple nodes. so maybe we need a top haproxy that combines multiple tfs on different nodes and access them via one haproxy or a hierarchy of haproxy