The growth process of `local_step` seems to have an issue.

[boyaom]

During the nerf training phase, the value of local_step is incremented by one separately in the run_cuda and train_step functions.
torch-ngp/utils.py#L858
torch-ngp/renderer.py#L939
Therefore, local_step will increase by 2 in each epoch, which leads to the step_counter updating only at even positions each time.