perf: stabilize streaming budget across sampling steps (+ log noise cleanup) #2519

Job	Run time
ubuntu-latest-cmake	6m 47s
ubuntu-latest-rocm (7.2.1, gfx908;gfx90a;gfx942;gfx1030;gfx1031;gfx1032;gfx1100;gfx1101;gfx1102;g...	5m 28s
windows-latest-cmake-hip	3m 6s
windows-latest-rocm	7m 14s
windows-latest-cmake (noavx, -DGGML_NATIVE=OFF -DGGML_AVX=OFF -DGGML_AVX2=OFF -DGGML_FMA=OFF -DSD...	3m 56s
ubuntu-latest-rocm (7.13.0, gfx906;gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gf...	4m 46s
ubuntu-latest-cmake-vulkan	9m 51s
windows-latest-cmake (avx2, -DGGML_NATIVE=OFF -DGGML_AVX2=ON -DSD_BUILD_SHARED_LIBS=ON)	4m 9s
macOS-latest-cmake	10m 20s
windows-latest-cmake (avx, -DGGML_NATIVE=OFF -DGGML_AVX=ON -DGGML_AVX2=OFF -DSD_BUILD_SHARED_LIBS...	3m 42s
windows-latest-cmake (cuda12, -DSD_CUDA=ON -DSD_BUILD_SHARED_LIBS=ON -DCMAKE_CUDA_ARCHITECTURES='...	1h 58m 42s
windows-latest-cmake (vulkan, -DSD_VULKAN=ON -DSD_BUILD_SHARED_LIBS=ON)	7m 23s
windows-latest-cmake (avx512, -DGGML_NATIVE=OFF -DGGML_AVX512=ON -DGGML_AVX=ON -DGGML_AVX2=ON -DS...	3m 41s
Build and push container images	-1s
release	0s
	3h 9m 4s

Provide feedback