This make the gpu version a little faster. More stuff to fix left.

when the 1msigmoid optimization is not registered.

can be unpickled properly even though Gemm has an 'inplace' property now.

Add parameter 'inplace' to Gemm Op.
Alias gemm_inplace to Gemm(inplace=True) and gemm_no_inplace to
Gemm(inplace=False).
Replace calls to 'gemm' in GemmOptimizer by calls to gemm_no_inplace.

If True, won't print the intermediate var that don't have name.

instead created setsubtensor function. Also added python and rst documentation.