Fix FFT execution on separate Julia task

[jipolanco]

This PR fixes the execution of an FFT plan on a different AMDGPU stream (or Julia task) from the one used to create the plan.

Example:

using AMDGPU
using AbstractFFTs

x = ROCArray(rand(Float32, 100, 200))

p = plan_rfft(x)
y = p * x

task = Threads.@spawn p * x   # this currently fails! (switches to a different stream)
z = fetch(task)

Array(z) == Array(y)  # should be true

This currently fails simply because the info variable is not defined in update_stream!.

[pxl-th]

Thanks!