changed default value of ConvOp.unroll_batch and ConvOp.unroll_kern to 4 as this is the best value for x86_64 computers.

dimensions (resulted in 3x speedup)

Please be careful when committing ;)

committing ! :(

mods to tensor/basic: added dimshuffle member function; changed smallest and largest to use switch() for 2-arg applications; added an additional error-check to numeric_grad.max_err

dx, dy != 1, 1

Now we return the same value with the same stride in the python version as the c version.