Pin CPU buffers when doing memory copies

[maleadt]

We should explore automatically pinning CPU memory buffers, or async memory copies are really executed synchronously:

using CUDA, LinearAlgebra

function expensive_computation(a,b,c)
    NVTX.@range "mul!" mul!(c, a, b)
    NVTX.@range "broadcast!" broadcast!(sin, c, c)
    Array(c)    # this cannot be executed asynchronously, because the destination memory can be paged out
end

# one "iteration", performing the above calculation twice in two tasks
# and comparing the output.
function iteration(a,b,c)
    results = Vector{Any}(undef, 2)
    NVTX.@range "iteration" @sync begin
        @async begin
            results[1] = NVTX.@range "run 1" expensive_computation(a,b,c)
        end
        @async begin
            results[2] = NVTX.@range "run 2" expensive_computation(a,b,c)
        end
    end
    results[1] == results[2]
end

function main(N=1024)
    a = CUDA.rand(N,N)
    b = CUDA.rand(N,N)
    c = CUDA.rand(N,N)
    synchronize()
    NVTX.@range "warmup" iteration(a,b,c)
    GC.gc(true)
    NVTX.@range "main" iteration(a,b,c)
end