when we are under memory pressure, or echo 3 > /proc/sys/vm/drop_caches,
kernel can send us many forget ops at the same time if we have lots of
inodes in memory. Running them all in goroutines can lead to even more
memory usage and eventually OOM.

Closes #21.

See #25.

MacOS apparently invokes StatFS on os.Lstat/os.Stat. Note that the
return fields are not actually populated here.

This was broken by golang/go@321c312d8246dec6889f5fe334b6193c320baf0e.

Go 1.8 renamed this function; cf. golang/go@87e48c5.

Necessary because our evil memclr should go away.

Benchmarks:

    benchmark                                          old ns/op     new ns/op     delta
    BenchmarkOutMessageReset/Single_buffer-12          4.94          4.95          +0.20%
    BenchmarkOutMessageReset/Many_buffers-12           10.4          12.0          +15.38%
    BenchmarkOutMessageGrowShrink/Single_buffer-12     42861         47666         +11.21%
    BenchmarkOutMessageGrowShrink/Many_buffers-12      101896        100626        -1.25%

    benchmark                                          old MB/s     new MB/s     speedup
    BenchmarkOutMessageReset/Single_buffer-12          4853.85      4851.64      1.00x
    BenchmarkOutMessageReset/Many_buffers-12           2311.82      2006.72      0.87x
    BenchmarkOutMessageGrowShrink/Single_buffer-12     24464.39     21998.19     0.90x
    BenchmarkOutMessageGrowShrink/Many_buffers-12      10290.60     10420.48     1.01x

Use int where appropriate, fix up documentation, and clarify. Delete the
implementation where it will soon need to change.