perf: add fast-math and vectorization nvcc/cxx build flags

- centralize CXX_FLAGS/NVCC_FLAGS in csrc/build.py as single source
- add --use_fast_math, --ptxas-options=-O3,-v, --extra-device-vectorization
- add -march=native -funroll-loops host flags
- setup.py reads shared cxx_flags/nvcc_flags from registry
- sync pure-C test build commands with new flags

csrc/tests/gqa_decode_test.cu

@@ -1,4 +1,10 @@
-// Pure-C test: nvcc -I csrc -arch=sm_89 csrc/tests/gqa_decode_test.cu -o test && ./test
+/*
+Pure-C test:
+nvcc -I csrc -arch=sm_89 -O3 \
+    --use_fast_math --ptxas-options=-O3 --extra-device-vectorization \
+    csrc/tests/gqa_decode_test.cu -o test && ./test
+*/
+
 #include <cstdio>
 #include <cstdlib>
 #include <cmath>

csrc/tests/gqa_prefill_test.cu

@@ -1,4 +1,10 @@
-// Pure-C test: nvcc -I csrc -arch=sm_89 csrc/tests/gqa_prefill_test.cu -o test && ./test
+/*
+Pure-C test:
+nvcc -I csrc -arch=sm_89 -O3 \
+    --use_fast_math --ptxas-options=-O3 --extra-device-vectorization \
+    csrc/tests/gqa_prefill_test.cu -o test && ./test
+*/
+
 #include <cstdio>
 #include <cstdlib>
 #include <cmath>