Use numbered subdirectories to avoid having too many files in one dir.
Add appropriate tests.

Also fix an issue where the mmap_open LRU cache could inappropriately
open a file twice because it was using the optional "newsize"
parameter as a key -- now lrucache can be given a slice object that
describes which arguments are important.

Now the goal is 128 MiB files, rather than a specific length.

We replace cStringIO with StringIO subclass that forces UTF-8
encoding, and explicitly convert commandlines to UTF-8 before
shlex.  These changes will only affect tests, not normal commandline
operation.

Spent way too long trying to track down a cryptic error that turned
out to be a CherryPy bug.  Now we catch this using a decorator in the
'extract' and 'intervals' generators that transforms exceptions that
trigger the bugs into one that does not.  fun!

Cleaned up OS-specific path handling in nilmdb, bulkdata.

This will make migrating to my own data storage engine easier.

Warns at runtime if a class's close() method wasn't called before the
object was destroyed.

There's some bug with the testing harness where placing e.g.
  from du import du
in nilmdb/utils/__init__.py doesn't quite work -- sometimes the
module "du" replaces the function "du".  Not exactly sure why;
we work around that by just renaming files so they don't match
the imported names directly.

Messes up extraction, since we random access for the timestamp binary
search.  In the future, maybe switching to multiple tables (one for
timestamp, one for compressed data) would be smart.

Adjust test expectations accordingly, since the number of intervals
they print out will now be smaller.

timeit.sh script works too!

Otherwise timestamps get truncated to 2 decimal places.

structure manually.