In prediction, avoid using broadcast, use ModelBroadcast

[yiheng]

In prediction, users may directly broadcast model to use it in each task without clone. This is bad because model has its own buffer which cannot be shared among threads. A better practice is that each task should use a model sharing the parameters but has its own buffer. BigDL provides a ModelBroadcast API for it.

We need to throw some warning if users fall into the first case. Current message is not helpful