nccl4py: report the loaded NCCL .so path and per-component versions

Resolve the actually-loaded library path via dladdr() on a resolved entry
point instead of scanning /proc/self/maps, so get_version()/show_versions()
report the .so the process is really using (e.g. a host-preloaded copy).

Signed-off-by: Xiakun Lu <xiakunl@nvidia.com>

Mirrored-from: 1acbe62d713c945a6b3e35a6bbfac3b499de1b3f