Warmup updates bug for LR < 1

[emilydinan]

Patch description
#4242 introduced a bug in which _is_warming_up returns True if the last LR is < 1. This will introduce a bug for all initial LRs != 1. In particular, for models with initial learning rate < 1, the model will be "warming up" forever, and the LR will remain constant after the warmup updates have finished rather than starting to decay according to the provided schedule.

Testing steps

parlai tm -t convai2 -m transformer/generator --lr-scheduler linear --warmup-updates 10 -lstep 1 -vstep 10000000 --max-lr-steps 100 --skip-generation True --warmup-rate 0.01 -lr 0.00001 --dict-file /tmp/test123.dict -mf /tmp/test1234dsfsdf5

I had to relax restrictions to get tests to pass. If we change self._number_training_updates < self.warmup_updates -->, self._number_training_updates <= self.warmup_updates, we hit the exact max LR, but don't quite anneal to zero. Will leave it to follow up PR (Jude) to test this more robustly

[stephenroller]

Ty

[emilydinan]

Tests pass locally

[stephenroller]

@klshuster and @jxmsML confirm this does NOT affect reduceonplataeu

[stephenroller]

@spencerp found this did NOT affect invsqrt