Faster reduction modulo L

Faster reduction modulo L

Replaced TweetNaCl's code by Barrett reduction.

- mod_l(), reduce(), and mul_add() use less stack.
- mod_l(), reduce(), and mul_add() are now much faster.
- Signature generation is noticeably faster (~7% on my Skylake laptop).
- Now I understand all the code. No more black boxes.

The change cost a total of 8 lines of code.