added optional argument (warn_type) to gradient.grad_sources_inputs to disable gradient type checking

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