#2913 Fix Qwen3.6 CUDA 12.0 build in debug half-bit print

Description

Building 0.1.45 with CUDA 12.0.140 fails in `Engine::verify_q8`: `__half_as_ushort` is device-only in that toolkit, but the diagnostic runs on the host.

Read the converted half's two bytes with `memcpy` and a size assertion. This only changes `STRATA_Q36_VDEBUG` reporting; inference arithmetic and defaults are unchanged.

Validated: full `strata` build with CUDA 12.0.140, GCC 12.4 and sm_61; `--version`, `--families`, and Qwen3.6 Q4 text generation on a Tesla P4. HIP was not compiled. Related older host-half failure: #2322.