fix(cuda): MTP + sm_89 compatibility for GCC 12 host compiler

[altifilmperisi]

Summary

• Disable -compress-mode flag — incompatible with GCC 12 host compiler (passed to GCC instead of nvcc)
• Guard FP8 __nv_fp8_e4m3 conversion behind __CUDA_ARCH__ >= 900 — FP8 PTX instructions require sm_90+, causing build failure on RTX 4060 Ti / Ada Lovelace (sm_89)

Why

Without these fixes, building with -DGGML_CUDA=ON fails on sm_89 GPUs (RTX 4060 Ti, 4070, 4080, 4090) when using GCC 12 as the CUDA host compiler:

1. gcc-12: error: unrecognized command-line option '-compress-mode=...'
2. ptxas: Feature 'cvt with .e4m3x2/.e5m2x2' requires .target sm_90 or higher

These changes allow llama-server to build successfully with CUDA support on Ada Lovelace GPUs while retaining full MTP (Multi-Token Prediction) support from upstream.

[TheTom]

Thanks. The __CUDA_ARCH__ >= 900 guard on the fp8 e4m3 path is the right call, sm_89 has to take the fallback. Builds clean on sm_121 here with no numeric change. Merged. Minor note for later: disabling -compress-mode globally fixes gcc12 but drops binary-size compression on every config, so a host-compiler version guard would be tidier down the line. Not blocking. Appreciate the fix.