RFE: Improve readTimeout handling

[pjdarton]

Some time ago, I introduced a readTimeout field.
I've since found that this doesn't go far enough...

• What we want is to have a nice short readTimeout normally so that docker API calls that should complete quickly will either succeed or fail quickly, thus preventing the plugin from locking up.
• However there are some long running operations which fail if we have a "nice short readTimeout", e.g. a docker pull operation doesn't return until it's complete and a docker image build can go quiet (so no new data to read) while it's processing lengthy "apt get" commands and similar.

It seems to me that DockerAPI needs to provide different connections with different timeouts for different operations.
...and if it's going to do that then we really need to ensure that we don't end up leaking connections (which we can do at present, see #615 and #617).

What I think we should do is to make DockerAPI cache connections using a Guava cache (or similar) that's set to throw stuff away (closing it first) if it hasn't been used for ages, and support having a normal "readTimeout", a "docker pull timeout" and maybe other timeouts too (TBD).

Note: Whilst this note is partly a "note to self" as a precursor to me getting on and doing it myself, I'd also appreciate hearing others' opinions on this.

[pjdarton]

Merged.
It'll be in the next release.