Fix integer overflow in ANN kmeans

[Nyrio]

CUDA block and thread indices are uint32, so the following operation results in an overflow before being cast to uint64:

uint64_t gid = threadIdx.x + (blockDim.x * blockIdx.x);

This PR fixes the error that @tfeher and myself have encountered while benchmarking large datasets:

Incorrect mesocluster size at 0. 625618 vs 625858

cc @achirkin

[tfeher]

Thanks for the fix Louis, LGTM!

[cjnolet]

@gpucibot merge