Implement break-down mode for cross entropy calculation

[penkoba]

This PR prevents the training process from falling down to the CPU mode, possibly happens in following case.

• large number of objects per image (detected and ground truth)
• large number of classes

When calculating cross entropy, the process requires a tensor with following size
nr_of_detected_objects * nr_of_ground_truth * nr_of_classes
which may explode in the above case and cause GPU OOM.

This PR implements break-down mode of cross entropy calculation, which requires only
nr_of_detected_objects * nr_of_classes
of tensor. This mode will be tried when the original code caused OOM, and still runs far faster than the CPU mode.

For testing, you can pull the following branch
https://github.com/penkoba/YOLOX/tree/break_down_cross_entropy_calc-test
and run test-train.sh, then you may see entering the break-down mode but still runs at acceptable speed.
Without this PR, the traning process falls into CPU mode very often and gets very slow.
(tested with RTX 2080 Ti)

[CLAassistant]

All committers have signed the CLA.