Fish S2 Pro AMD/HIP AR optimization on RX 7900 XTX

[reezex0-ux]

Profiled Fish S2 Pro AR on RX 7900 XTX / ROCm 7.2.1. GPU-side sampling reduced warm mean 4.291s -> 3.667s (~14.5%); experimental chained Fast-AR reached 3.557s (~17.1%). Documented test WAV was bit-identical. Patch, benchmark method, and notes: https://github.com/reezex0-ux/fish-s2-amd-hip . I checked current main: Fish code has been reorganized, so this patch should not be applied blindly, but the CPU sampling/readback pattern remains. Would there be interest in upstreaming the useful pieces as small current-main PRs (sampling first, chained Fast-AR kept separate)? Only RX 7900 XTX has been benchmarked so far; cross-testing is welcome.

[0xShug0]

@reezex0-ux Thank you for the optimizations! Could you submit a PR with the changes that are additive and can be scoped to HIP first?

[IIIIIllllIIIIIlllll]

Thanks! I can do some test this weekend :)