… benchmark helper functions; and actual benchmark tables, runnable, which compare several storage implementations.

Mostly deredundantizing, but made a few improvements as well.

…default configuration of go-ipfs.

Include links to where the relevant ipfs code can be found.
(It's not all in very obvious places, and not all in the go-ipfs repo.)

Extracted some of the b32 encoder to a package function...
but, although this does avoid some repeated initialization,
it was all CPU stuff.  Benchmarks indicate that it made no real
difference.  So it's mostly a stylistic change.

One level with fmt.Sprintf is enough for what we need,
and is also easier to follow.

... which is actually important, not just stylistic: this also makes
the benchmarks much more correct, because we were previously
doing the cleanup work on the clock!  Now the cleanup is done by
the test system, off the clock.

This fixes results to actually be, um, correct.  *facepalm*.

The storage APIs now require this and caution of potential resource
leakage otherwise.

…ng parameters that are in fact "default" when comparing these.

…an io.ReadCloser instead of just io.Reader.

This seems more likely to encourge good code that doesn't leak
resources than just admonishing people to check via a comment.

On the downside, sometimes this requires implementations to wrap stuff
in no-op closers.  The trade is probably worth it.

The code in LinkSystem which uses these things still doesn't change.
It's still using the `linking.Block*Opener` functions, which still
return just io.Reader.  And we're considerably less free to change
those, because there are already quite a few implementations of those
types in the wild, and we don't want to break them.

(I think our policy is roughly "support the last two significant go
versions", and by that, we might technically be allowed to use this,
but, eh; there's minimal cost to being conservative today.)