Only write the modified tablets to file system.

[jetfuel]

Move some functions definitions from headers to cc files.

The approach is to track what tag is modified in the storage during training. Only process the modified tablets at the PersistToDisk.

This should cover the first step in #284.
We should also look into if we can only call persistToDisk once at the end of training. Protobuf always serialize all data to the filesystem and that's a huge waste.

[Superjomn]

MarkTableModified use a set and that takes an overhead if AddRecord is triggered frequently(e.g. in a scalar), and theoretically, it should be placed in WRITE_GUARD so that modifying a record will also mark the tablet "modified". But currently, this implementation is ok.

There is a trick that can be faster, assigning each tablet an internal member id_ (continuously ascending from zero), and make modified_tablet_set_ a vector, the marking logic like this

vector<bool> modified_tablet_set_; // each time add a tablet will append a true
Record Tablet::AddRecord() {
  parent()->MarkTabletModified(id_);
  // ...
}
inline void Storage::MarkTabletModified(int tagid) {
  modified_tablet_set_[tagid] = true;
}
void Storage::PersistToDisk(const std::string& dir) ... // clear states

[jetfuel]

I spent some time on converting to vector, but encounter some errors when adding records. Will put this on hold for a bit. :/