Standalone execution of fused kernels

[ezyang]

I wanted to use the JIT generated fused kernel from C++ (to help figure out how much is autograd overhead), but there was no way to easily do so. It would be good if there were.

[zdevito]

The jit tests have some examples of using the fusion engine from C++. There are also version that separate the cache-checking from the execution:

https://github.com/pytorch/pytorch/blob/master/torch/csrc/jit/test_jit.cpp#L99