Fix QNX fill_workspace status

[fallintoplace]

Summary

fill_workspace() currently skips the CUDA driver memset calls on QNX, but leaves result initialized to CUDA_SUCCESS. That means a non-empty fill can be reported as successful even though the direct-CUDA path did not write the workspace.

This changes the non-adapter QNX path to return Status::kErrorInternal instead of silently succeeding. The CUDA host adapter path is left unchanged, so QNX builds that enable and pass an adapter can still use that route.

I also fixed the nearby cuGetErrorString() call to pass the address of a const char*, instead of passing a null const char**.

Validation

• git diff --check
• source-level check that the QNX direct path returns an error and cuGetErrorString() receives &error_string