Fix correctness in cuda_mapreduce

[Sbozzolo]

cuda_mapreduce was not working correctly with certain spaces.

Why was this happening?

I added a comment to describe the algorithm in the commit.

In a nutshell, the algorithm was not taking into account the fact that the final block is not completely filled with points to process. Therefore, the reduction included some elements that did not contain real points (but the value 0).

Closes #2097

[sriharshakandala]

One option is to use binary-op appropriate initialization. For example,

function _init_val_for_reduction(f::Function, ::Type{T}) where {T}
    f == min && return typemax(T)
    f == max && return typemin(T)
    return T(0)
end

with

reduction[tidx] = _init_val_for_reduction(op, T)

[Sbozzolo]

One option is to use binary-op appropriate initialization. For example,

function _init_val_for_reduction(f::Function, ::Type{T}) where {T}
    f == min && return typemax(T)
    f == max && return typemin(T)
    return T(0)
end

with

reduction[tidx] = _init_val_for_reduction(op, T)

This would require defining the init value for every function, which doesn't seem optimal.

Is there any issue with the fix in this PR?

[sriharshakandala]

Each thread loads multiple data points in shmem!