Add host-side bfloat16 helpers for platforms without std::bfloat16_t

[thomthehound]

Under the C++23 standard, std::bfloat16_t support is optional. The current MSVC toolchain has chosen not to exercise that option, while the Linux toolchain does. Therefore, host-side example code that manipulates bfloat16 data cannot use std::bfloat16_t safely when compiled with MSVC, as is the intended route on native Windows.

This patch adds test_utils::bfloat16_t (and related helpers) to test_utils.h. When the standard type is available, test_utils::bfloat16_t resolves directly to std::bfloat16_t, preserving native execution. Otherwise, it uses std::uint16_t storage and performs host-side conversion/math through float (rounding back to 16 bits by the nearest-even method).

In the trivial case, the type selection works as so:

#if defined(__STDCPP_BFLOAT16_T__)
using bfloat16_t = std::bfloat16_t;
#else
using bfloat16_t = std::uint16_t;
#endif

These changes are host-side only; device code is unaffected. The examples have been updated to use test_utils::bfloat16_t consistently for host-side bfloat16 data (including those that are not obviously Windows-portable) so there are not two competing vocabularies.

This workaround is not intended to permanently solve host-side bfloat16 portability for all user code. It is intended to draw attention to this limitation, and to prepare the programming examples for cross-platform conversion. In the future, it may be reasonable to provide a better API boundary for this, but doing so would create a new public convention and is therefore a separate consideration.