Improve Arbiter implementation and QoR

- Remove power-of-2 restriction on count
- Avoid redundancy between lockIdx and chosen signals, reducing area
- Infer locked from lockCount, rather than maintaining separate state
- Reduce FIRRTL node count in several places
- Special-case simple Arbiter, which avoids serializing the computation
  of io.chosen with io.out.valid/bits.  This lops off a few FO4 delays
  at no area cost, and actually reduces area if io.chosen isn't used.