Hyper-optimized megakernel-hybrid transformer inference
We split the 14 given test cases into three families. The first two families dispatch to one bespoke megakernel each. The last family is handled by one of three highly optimized hybrid strategies.
