Add simple mode to Eval

[apaszke]

For first backward, the Eval wraps only a single function, and we can save a lot of allocations and cycles by skipping initialization of all data structures we need to perform the graphs search for a "general case".

This is one approach to do it, but now I think it might be even simpler to just provide another replaceSubgraph overload, that would do no graph traversal and would be used only by forward functions (which is a bit awkward, because we don't always know what C++ functions are forward/backward; I'm not sure if adding yet another virtual method is a good idea).

Word language model run times:

	forward	backward
No JIT	40ms	72ms
JIT	49ms	105ms
JIT + simple Eval	37ms	75ms

[apaszke]

The most straightforward changes that remove some STL containers from Eval::getSubgraph cause the runtime to drop to 39ms for forward and 85ms for backward. And I still have things I can optimize, so I expect it to get even better!

Maybe simple Eval helps so much just because it doesn't perform any allocations for the graph traversal. We've been using 8 STL containers, and most of them were sets (which allocate a separate node for each pointer they hold :blush:).

[apaszke]

I really don't think the DFS will be causing us any problems. We're doing it every time it runs the engine, but it always shows up to have a marginal overhead in the profiler (especially that we DFS over small hot parts of the graph that are very likely to be in the cache).

[ezyang]

LGTM, but tests need fixing.

[zdevito]

Seems good to me. I would need more background in the current system to have more substantive comments, but go ahead and merge once tests pass.