Fix device-side reduce launch with large inputs

[gevtushenko]

When there are more than one tile of input items, we launch two kernels for reduction. During the second kernel launch, we accumulate partial aggregates from thread blocks. In this case, int reduce_grid_size is passed to OffsetT num_items kernel parameter. Thrust triple_chevron is not type safe when used on device. These facts lead to a mismatch of data size we put into a param buffer (4 bytes) and that the kernel reads from the buffer (8 bytes). This seems to be a bug that was long present in thrust, that uses dispatch layer directly with offset types different from int32_t. In the case of host-side usage, type conversion takes place, leading to normal execution.

[elstehle]

Interesting finding 👍

Seems a bit concerning that device-side triple_chevron::do_it isn't type safe. Could imagine that one can easily instantiate a kernel template for different parameter types then are eventually getting passed to the kernel invocation.