[SYCL][NVPTX][AMDGCN] Don't pollute global namespace in cmath wrappers (#19620)

Before this change we defined cmath overloads like `float fabs(float)`,
`float fabsf(float)`, `double fabs(double)` in the global namespace.
Normally, standard libraries only define the C functions there (`float
fabsf(float)`, `double fabs(double)`) when including `<cmath>`.
By defining all overloads in the global namespace, we were failing to
compile the following code in SYCL mode that works under normal C++
compilation (without offloading):
```cpp
#include <cmath>

void f(long double x) {
    // This call is ambiguous, because both `double fabs(double)` and
    // `float fabs(float)` are defined in the global namespace, but not
    // `long double fabs(long double)`.
    auto y = fabs(x);
    // Note that this works both before and after this change:
    auto z = std::fabs(x);
}
```

This was spotted in Intel internal code, and while it could be argued
that this is a bug in that code, I believe we should only break
compatibility when it is necessary.

This change moves the overloads and templates that are C++-only into the
`std` namespace, while keeping the C functions in the global namespace.
This is the same behavior as libstdc++ and libc++.

Author: Maetveis