refactor: root kernel includes at csrc/kernels

- add the kernels directory to CMake target include paths and drop all ../-relative includes in kernel sources
- reference shared primitives as common/*.cuh and the fp8 type header as fp8/common.h
- update standalone test nvcc commands in file headers and cuda_kernels.md to -I csrc/kernels

csrc/kernels/attention/mma_utils.cuh

@@ -3,8 +3,8 @@
 #include <cuda_fp16.h>
 #include <cuda_runtime.h>
 
-#include "../common/cp_async.cuh"
-#include "../common/mma.cuh"
+#include "common/cp_async.cuh"
+#include "common/mma.cuh"
 
 // Predicated cp.async (4-operand form) requires CUDA 11.2+.
 // bf16 mma.sync requires sm_80+ (guarded at build time by ASTRAI_NO_MMA).