Fix target specific option mismatch during OpenMP loop inlining under GCC 16

[Soddentrough]

Description

On bleeding-edge Linux distributions (like Fedora 44) running GCC 16.1.1+, compilation of `csrc/cpu_ops.cpp` fails when using OpenMP and AVX-512 flags. 

GCC 16 outlines OpenMP loop bodies into helper functions that do not always carry the parent's target attributes. When `std::fill` (marked `always_inline`) is called inside this loop, GCC reports a `target specific option mismatch` and aborts.

This PR replaces the single call to `std::fill` with a standard `for` loop, eliminating the compiler's function inlining target mismatch.

### Error log
```
FAILED: CMakeFiles/bitsandbytes.dir/csrc/cpu_ops.cpp.o
/usr/include/c++/16/bits/stl_algobase.h:1001:5: error: inlining failed in call to ‘always_inline’ ‘void std::fill(_ForwardIterator, _ForwardIterator, const _Tp&) [with _ForwardIterator = unsigned char*; _Tp = int]’: target specific option mismatch
 1001 |     fill(_ForwardIterator __first, _ForwardIterator __last, const _Tp& __value)
      |     ^~~~
/home/naoki/src/bitsandbytes/csrc/cpu_ops.cpp:326:22: note: called from here
  326 |             std::fill(out + block_start, out + block_end, 0);
      |             ~~~~~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
```