[FEA] Support eval mode in dynamicemb

[jiashuy]

Is your feature request related to a problem? Please describe.
Under eval mode in pytorch, we shouldn't change the weight.
However, HKV may evict keys in the dynamicemb's forward, causing weigth changes.

Describe the solution you'd like
Identify the status(model.train() / model.eval()) and select different code paths in dynamicemb.
For eval mode, using find and not find_or_insert to lookup the HKV.

[shijieliu]

in inference, the missed key can be initialized as 0. So we dont need unique in eval

[shijieliu]

close as we dont plan to support eval for dynamicemb. We will use nvembedding for the eval.