atomic.add operation leads to incorrect results

[diptorupd]

A test micro-benchmark implemented in dpbench (https://github.com/IntelPython/dpbench/blob/main/dpbench/benchmarks/l2_distance/l2_distance_numba_dpex_k.py) produces incorrect results. The most likely culprit is the usage of atomic.add.

The generated code for numba-dpex's atomic add function needs to be evaluated v/s dpcpp.

[diptorupd]

@mingjie-intel this might be related to #778. Maybe the llvm-spirv we are using is too archaic and generates incorrect code.

[diptorupd]

Closing the issue as the benchmark no longer exists and we cannot reproduce the problem. In any case, the fix in #1103 will fix the original issue.