Eliminate Sendbuffer

[Alex2804]

Implement a version of the shuffle that sends directly tuple-by-tuple using nvshmem_put.
For this, we have to have a thread-level histogram to calculate thread-level write offsets on the destinaiton PEs.
We also need to make sure that the input data is in symmetric device memory. We can register it as symm. mem. if possible or allocate symm. dev. mem. and copy the data over.

[Jonas-Wessner]

Not as relevant since micro benchmarks show that send buffers are required to get to good performance, since tuple-wise sending would result in too many small put calls.
Nevertheless, feature is kind of done but has a small bug.