Fix atomic_ref scope when multiple blocks are updating the same output

[vuule]

Description

in a few places, thread_scope_block is used even where there threads from multiple blocks update the same location. This PR changes these to thread_scope_device to avoid UB with sufficiently large inputs.

Have not ran benchmarks to evaluate the impact.

Checklist

• I am familiar with the Contributing Guidelines.
• New or existing tests cover these changes.
• The documentation is up to date with these changes.

[PointKernel]

LGTM!

Thanks for the fix!

[vuule]

/merge