specialization.  It inserts a new op called PseudoGemm.  This Op has the same
signature as Gemm but does not work inplace.   Another optimization comes
later in the pipeline and swaps PseudoGemm for Gemm

An Env feature persists throughout the life of the env, so many optimizations
can take advantage of the Shape analysis done by ShapeFeature.

exact output from the rngs, the float32-casted sampling bounds mess things up.

phase.

graphs.

of infer_shape

optimizer.

Also, moved MakeVector to tensor.opt since it is only used internally by
ShapeOpt.

that is scheduled to run first, before the first merge.

first when sorted.

Canonicalizer's get_num_denum function... this might break tests, but it was
essential for getting rid of numerical instabilities in RBM free energy grad.

internally, to print results, apply nodes, and functions (with multiple
outputs)

indentation in big graphs... still a bit ugly though.