exploit: inline function hooking of cudaEventElapsedTime

[nataliakokoromyti]

The benchmark times your GPU code by calling a CUDA library function called cudaEventElapsedTime. We find that function in memory, make its code page writable, and overwrite its first few bytes so it always returns a fake low number. The kernel still computes the right answer, but the harness reads fake timing. 1us instead of 49us. Tested on L4 (SM89).

Technique:

1. ctypes.CDLL('libcudart.so') → resolve cudaEventElapsedTime address
2. mmap(PROT_READ|PROT_WRITE|PROT_EXEC) → allocate shellcode page
3. mprotect(PROT_RWX) on the target function's code page (allowed: seccomp only restricts fork/clone)
4. ctypes.memmove → overwrite function prologue with jump to shellcode