GPU compilation failure with pow.hpp from boost::multiprecision

I have some C++ code that compiled just fine with `nvcc` from CUDA 12.6.1 and Boost 1.84.0. After updating to Boost 1.88.0, which includes the GPU math changes from Boost 1.87.0, I'm now seeing the following error message:

```
/usr/include/boost/multiprecision/detail/functions/pow.hpp(789): error: namespace "boost::math" has no member "rounding_error"
catch(const boost::math::rounding_error&)
^
```

If I'm understanding Boost's macro and include header structure correctly, boost/multiprecision/detail/functions/pow.hpp is inlined with boost/multiprecision/detail/default_ops.hpp. pow.hpp is always included no matter what in default_ops.hpp. `boost::math::rounding_error` is defined in boost/math/policies/error_handling.hpp, which is only included by default_ops.hpp if `BOOST_MP_MATH_AVAILABLE` is defined. This is the same macro that's used for guarding `BOOST_MP_CATCH(const boost::math::rounding_error&) {}` in pow.hpp (the same code not compiling), so all seems good. But, `boost::math::rounding_error` is only defined when `BOOST_MATH_NO_EXCEPTIONS` is NOT defined. And, when compiling with `nvcc`, `BOOST_MATH_NO_EXCEPTIONS` DOES get defined, according to boost/math/tools/config.hpp, because we've got `__CUDACC__` defined by simply using `nvcc`.

Should we be adjusting pow.hpp, and possibly any other cases like this in either `boost::math` or `boost::multiprecision`, to have something like this instead?

```
#if defined(BOOST_MP_MATH_AVAILABLE) && !defined(BOOST_MATH_NO_EXCEPTIONS)
BOOST_MP_CATCH(const boost::math::rounding_error&)
{ /* Fallthrough */
}
#endif
```

That way we only define this catch condition when we have exception support and multiprecision math is available?