- use slice ops for convertWords instead of lo/hi boundaries
- always compute leading zeroes (simplifies logic significantly),
  but remove them once, at the end (since leafSize is small, the
  worst-case scenario is not adding significant overhead)
- various comment cleanups (specifically, replaced direct -> iterative,
  and indirect -> recursive)
- slightly faster overall for -bench=String

(This CL incorporates the changes re: my comments to CL 5418047
https://golang.org/cl/5418047/ )

benchmark                          old ns/op    new ns/op    delta
big.BenchmarkString10Base2               519          527   +1.54%
big.BenchmarkString100Base2             2279         2158   -5.31%
big.BenchmarkString1000Base2           18475        17323   -6.24%
big.BenchmarkString10000Base2         178248       166219   -6.75%
big.BenchmarkString100000Base2       1548494      1431587   -7.55%
big.BenchmarkString10Base8               415          422   +1.69%
big.BenchmarkString100Base8             1025          978   -4.59%
big.BenchmarkString1000Base8            6822         6428   -5.78%
big.BenchmarkString10000Base8          64598        61065   -5.47%
big.BenchmarkString100000Base8        593788       549150   -7.52%
big.BenchmarkString10Base10              654          645   -1.38%
big.BenchmarkString100Base10            1863         1835   -1.50%
big.BenchmarkString1000Base10          12099        11981   -0.98%
big.BenchmarkString10000Base10         57601        56888   -1.24%
big.BenchmarkString100000Base10     20123120     19827890   -1.47%
big.BenchmarkString10Base16              358          362   +1.12%
big.BenchmarkString100Base16             815          776   -4.79%
big.BenchmarkString1000Base16           4710         4421   -6.14%
big.BenchmarkString10000Base16         43938        40968   -6.76%
big.BenchmarkString100000Base16       406307       373930   -7.97%

R=michael.jones, mtj
CC=golang-dev
https://golang.org/cl/5432090