fix: restore set_recent_kernel(2) in rate_limiter() to match original behavior

Restore the unconditional set_recent_kernel(2) call that was removed
in the rate_limiter optimization. The write has negligible cost (~100-
200ns cache line store) compared to the other savings in this function,
and removing it changes observable shared memory state which could
affect external tooling or future features.

The call is placed after the cached sm_limit/util_switch fast-exit,
matching the original position relative to the get_recent_kernel()
guard. All other optimizations (cached limits, removed duplicate
sm_limit call, reduced sleep) are preserved.

Signed-off-by nishitnshah <nishshah@linkedin.com>

Signed-off-by: Maverick123123 <yuming.wu@dynamia.ai>

Author: nishitnshah

src/multiprocess/multiprocess_utilization_watcher.c

@@ -54,6 +54,7 @@ void rate_limiter(int grids, int blocks) {
   while (get_recent_kernel() < 0) {
     usleep(1000);
   }
+  set_recent_kernel(2);
 
   LOG_DEBUG("grid: %d, blocks: %d", grids, blocks);
   LOG_DEBUG("launch kernel %ld, curr core: %ld", kernel_size, g_cur_cuda_cores);