Write() can safely use uint32 loads when input is aligned.
Also add test and benchmarks for unaligned writes.

Benchmark result obtained by Dave Cheney on ARMv5TE @ 1.2GHz:
benchmark                       old ns/op    new ns/op    delta
BenchmarkHash8Bytes                  4104         3417  -16.74%
BenchmarkHash1K                     22061        11208  -49.20%
BenchmarkHash8K                    146630        65148  -55.57%
BenchmarkHash8BytesUnaligned         4128         3436  -16.76%
BenchmarkHash1KUnaligned            22054        21473   -2.63%
BenchmarkHash8KUnaligned           146658       146909   +0.17%

benchmark                        old MB/s     new MB/s  speedup
BenchmarkHash8Bytes                  1.95         2.34    1.20x
BenchmarkHash1K                     46.42        91.36    1.97x
BenchmarkHash8K                     55.87       125.74    2.25x
BenchmarkHash8BytesUnaligned         1.94         2.33    1.20x
BenchmarkHash1KUnaligned            46.43        47.69    1.03x
BenchmarkHash8KUnaligned            55.86        55.76    1.00x

R=golang-dev, dave, bradfitz
CC=golang-dev
https://golang.org/cl/6782072