perf flame graphs at perf_event_paranoid=2 with io_uring servers

Without root at perf_event_paranoid=2, perf record -g is denied. Use -e cpu-clock:u --call-graph dwarf --user-callchains, which works without root. If the target process sets up an io_uring ring, launching it under perf can fail with ENOMEM because perf's buffers plus the ring exceed the default 1 MiB RLIMIT_MEMLOCK. Start the process normally and attach with perf record -p <pid> instead. Also, when profiling or stracing a server started through a wrapper, signal the real server pid, not the wrapper, or the report comes back empty.