Faster att mask gen

[JohnGiorgi]

Overview

This PR is a combination of a few things, just rolled them into one as we are rushing toward the ICLR deadline.

1. Replaces list comprehensions used to generate attention masks with torch.where. This is many times faster.
2. Remove the transformer layer. It was accounting for almost no performance (and actually hurting performance in some cases).
3. Clean up the optimizer initialization code. It is now better written to prevent weight decay being applied to normalization layers.

Something funky happen and I had to push the same commits twice.

[coveralls]

Pull Request Test Coverage Report for Build 494

• 0 of 0 changed or added relevant lines in 0 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage remained the same at ?%

---

Totals
Change from base Build 490:	0.0%
Covered Lines:
Relevant Lines:	0

---

💛 - Coveralls

[coveralls]

Pull Request Test Coverage Report for Build 496

• 13 of 14 (92.86%) changed or added relevant lines in 3 files are covered.
• No unchanged relevant lines lost coverage.
• Overall coverage remained the same at 88.084%

---

Changes Missing Coverage	Covered Lines	Changed/Added Lines	%
saber/utils/bert_utils.py	10	11	90.91%
<!--	Total:	13	14	92.86%	-->

Totals
Change from base Build 490:	0.0%
Covered Lines:	1434
Relevant Lines:	1628

---

💛 - Coveralls