Summary

This fixes the CUDA initialize_pencil example kernel used by the basic usage examples.

The kernel computes a flattened pencil element index as l, but wrote initialized values through data[i]. That means threads with the same first-dimension coordinate race on the same buffer slots, while most of the pencil allocation is left untouched.

This PR updates the device initialization path to store through data[l], matching the host-side flattened loop already shown in the example. It also tightens the guard to l >= pinfo.size so the extra threads from the rounded-up launch do not write past the valid pencil range. The documentation snippet is updated to match the source examples.

Validation

• git diff --check
• Searched the touched example/docs scope to confirm the old data[i] store and l > pinfo.size guard are gone

I could not build or run the CUDA/MPI example locally because this machine does not have nvcc or mpicxx installed.