Fix cub::DeviceHistogram::MultiHistogramEven() 1-Dimensional overload, when OffsetT is not an int.

[nv-dlasalle]

When using gcc=7.5.0 and OffsetT=int64_t, the following template errors occur:

/usr/local/cuda/include/cub/device/device_histogram.cuh(144): error: no instance of overloaded function "dgl::cub::DeviceHistogram::MultiHistogramEven" matches the argument list
            argument types are: (void *, size_t, int32_t *, unsigned long long *[1], int [1], int32_t [1], int32_t [1], int64_t, int, unsigned long, cudaStream_t, __nv_bool)
          detected during:
            instantiation of "cudaError_t dgl::cub::DeviceHistogram::HistogramEven(void *, size_t &, SampleIteratorT, CounterT *, int, LevelT, LevelT, OffsetT, cudaStream_t, __nv_bool) [with SampleIteratorT=int32_t *, CounterT=unsigned long long, LevelT=int32_t, OffsetT=int64_t]"

This resolves this, by casting the 0 to an OffsetT, so as to ensure num_rows and num_row_samples are of the same type.

[alliepiper]

Thanks for the patch! This LGTM.

Would you mind updating this PR to fix the same issue on line 153 of this file, too?

[nv-dlasalle]

@allisonvacanti Updated line 153 as well--which is actually the one where I encountered the issue. Thanks for catching that!

[alliepiper]

That was fast -- thanks! I'll get this in for the next release.

[alliepiper]

DVS CL: 29946795
gpuCI: NVIDIA/thrust#1428

[alliepiper]

Tests look good -- I'll merge this soon.