Fix Dynamic BN for inefficient memory use

[yuikosakuma1]

This PR addresses the following two issues:

1. Dynamic BN operation contained undesirable graph construction that caused linear increase in memory use each iteration.
   This PR fixes this bug by modifying DynamicBatchNorm.

You can check this by adding the following codes in nnabla_nas/runned/searcher/ofa.py, L253 or anywhere in the training loop.

from nnabla_ext.cuda.init import print_memory_cache_map
if self.comm.rank == 0:
    print_memory_cache_map()

This gives something like cache_map(device_id: 0, mem_type: small, used: 1.86GB, free: 10.55MB):
You can check used and see this doesn't increase after several iterations.
This can be checked by running elastic expand_ratio for any OFA model. For example,

python main.py experiment=classification/ofa/ofa_mbv3/imagenet_search_expand_phase2

2. clear_memory_cache() was removed because this is not necessary and may slow the training.

[yuikosakuma1]

@hyingho Please merge this PR.