#166 build: replace CUDART_INF_F with INFINITY for HIP-only toolchains (fixes #157)

Description

`qsa_prompt_attn.cu` is already refused at runtime on AMD, but it still compiled with the CUDA-only `<math_constants.h>` header for `CUDART_INF_F`. Replace it with the portable `INFINITY` macro from `<cmath>` so the HIP build succeeds without a CUDA toolkit installed.

Fixes #157.