[bug fix] Torch Classifier agent should call self.backward(loss)

[dexterju27]

Patch description
The current torch classifier agent uses loss.backward instead of self.backward(loss). A lot of optimizer behaviors depend on the backward function. While the backward function was ignored in the previous version.
For example:

1. In SafeFP16Optimizer, the fp32 gradient will be zeros as no synchronizing flag is set. The model doesn't train.

   ParlAI/parlai/utils/fp16.py

   Line 200 in 66c71e0

   if self._needs_sync:

   .
2. In MemoryEfficientFP16Optimizer the loss loss_scale multiplier is omitted. We should improve performance with this fix.
   

   ParlAI/parlai/utils/fp16.py

   Line 521 in 66c71e0

   loss = loss * self.scaler.loss_scale

Testing steps
Train any classifier agent. Verified with a classifier agent training with SafeFP16Optimizer.
In our case, we retrained a safety classifier and compared its performance with the model in the zoo.
FP32 matched the performance, and FP16 performance worse. The performance shifted a bit between the tasks, we are comparing the average class_notok__ f1, it is the validation metric used for all the trainings.