Allow flan-t5 models in ParlAI with fp16 improvment

[dexterju27]

Patch description
The current setting does allow flan-t5 model to run under parlai hugging face agent seamlessly. Adding Flan-t5 options in the agent.

Testing steps

parlai train_model -t convai2  -dt train --batchsize 1  --fp16 True  --gradient-clip 1.0 --label-truncate 256 --text-truncate 512 --log-every-n-secs 30 --lr-scheduler reduceonplateau --max-train-time 169344.0 --model hugging_face/t5 --save-after-valid True --skip-generation True --optimizer adam -lr 1e-05 -veps 0.15 -vmm min -vmt ppl -vp 15 -tblog True --t5-model-arch google/flan-t5-large  --t5-model-parallel True --dict-tokenizer gpt2

[klshuster]

thanks!