changing variable name for inference start time for proper latency calculation when using --infer_time

[realbytecode]

sec_per_example is calculated from different start time than the experiment start and logs wrong value. Really short sec_per_example was also raised in these issues #1464 and #1468.

steps required to reproduce the bug

python3 test.py --cfg_file cfgs/nuscenes_models/cbgs_dyn_pp_centerpoint.yaml --ckpt ../checkpoints/cbgs_pp_centerpoint_nds6070.pth --infer_time is showing 2023-10-03 10:40:45,533 INFO *************** EPOCH 6070 EVALUATION ***************** eval: 100%|█████████████████████████████████████████████████████████████████████████████████████████| 162/162 [00:17<00:00, 9.33it/s, infer_time=58.17(101.47), recall_0.3=(0, 4375) / 5519] 2023-10-03 10:41:02,902 INFO *************** Performance of EPOCH 6070 ***************** 2023-10-03 10:41:02,902 INFO Generate label finished(sec_per_example: 0.0005 second).

This fix solves and shows the correct sec_per_example value. 2023-10-03 10:42:12,475 INFO *************** EPOCH 6070 EVALUATION ***************** eval: 100%|█████████████████████████████████████████████████████████████████████████████████████████| 162/162 [00:17<00:00, 9.22it/s, infer_time=53.93(101.71), recall_0.3=(0, 4375) / 5519] 2023-10-03 10:42:30,057 INFO *************** Performance of EPOCH 6070 ***************** 2023-10-03 10:42:30,057 INFO Generate label finished(sec_per_example: 0.1085 second).