CUDA-dyn dispatch: allow source ops to write to output buffer directly

[0ax1]

For scalar ops we can already write to either shared memory or to global memory (in case it's the last op).

In order to bypass unnecessary intermediate writes to shared memory, source ops should be able to write to the output buf directly. This can lead to significant perf improvements. E.g. CUDA runend decoding jumps from ~1500GiB/s to 3000GiB/s on a GH200.