This should help older GPUs run at all and newer GPUs fit more blocks
on one SM.

With this change the code is cc 2.0+ compatible.  But it will only be
fast on cc 3.0+ cards (due to atomicAdd).

Also some small kernel speedups elsewhere.

Also address some other issues that came up in code review.

Remove the beta parameter since it's always 1 anyway.

a dependecy on scikits.cuda and pycuda.

Update docs to reflect batches and add some fallback code to add batches of 1 to non-batched version.

Some improvements to the gradient, but it's still transposed and partially wrong.