gradient clipping by global norm

[ngc92]

one new kernel that calculates the overall norm of the gradient, and updates to the adam kernel.
Still TODO:

• clip value is hardcoded at function call site
• error handling for broken gradients would fit nicely into this change, but it currently only very rudimentary
• there is a debug print statement there. This debug print is in the kernel because we never move gradient norms to the host, but really, it should be removed once the code is ready for merge.
• python reference needs to be updated