Entropy pool

Code largely copied from the buffered-crypto branch.

See the README.md diff for a benchmark comparison. Full new benchmarks:

BenchmarkRead32-4             	10000000	       207 ns/op	 153.94 MB/s
BenchmarkRead512K-4           	     500	   3230813 ns/op	 158.47 MB/s
BenchmarkRead4Threads-4       	 2000000	       945 ns/op	 135.32 MB/s
BenchmarkRead4Threads512k-4   	     100	  14697556 ns/op	 139.34 MB/s
BenchmarkRead64Threads-4      	  100000	     14889 ns/op	 137.55 MB/s
BenchmarkRead64Threads64k-4   	       5	 235147804 ns/op	 139.35 MB/s
BenchmarkReadCrypto32-4       	  500000	      2697 ns/op	  11.86 MB/s
BenchmarkReadCrypto512K-4     	      50	  32161930 ns/op	  15.92 MB/s

In short, the throughput stays roughly constant even with 64 consumers, which is cool.