[LTC] Map a graph to a graph executor instance
Currently a new graph executor is created every time ExecuteComputation
happens. Instead we should use the same graph executor for the same
graph.
Roughly this cuts a single relu-op iteration execution time by half.
