* Add a custom softmax implementation. * Add softmaxlastdim to the benchmarks. * And add a test. * Support more dtypes. * Polish the code. * Use the slow implementation on cuda. * Add a todo for the cuda kernel.