[ML] Add timeouts to named pipe connections

[droberts195]

This PR adds timeouts to the named pipe connections of the
autodetect, normalize and data_frame_analyzer processes.
(The controller process already had a different mechanism,
tied to the ES JVM lifetime.)

Before this change if the ES JVM didn't connect to one of
the named pipes of an autodetect, normalize or
data_frame_analyzer process then it would hang forever.

After this change if nothing connects to the other end of
a named pipe that the C++ process has created within the
timeout period then it will delete the named pipe and exit.

The timeout will be supplied on the command line by a
followup Java PR. In the meantime, (unreleased snapshot)
builds that have the C++ change but not the Java change will
use a timeout of 5 minutes, which is better than nothing and
highly unlikely to be less than the configured process
connect timeout in the ES settings.

Relates #1504

[droberts195]

I decided to stick with TTime here rather than switch to std::chrono::duration. The reason is that all the existing command line argument parsing is currently using TTime and taking whole numbers of seconds, and it would be inconsistent if the new command line argument tried to do something clever with generic durations.

I think at some point we do need to move over to the std::chrono stuff throughout the codebase. This is obviously a huge project but unless we do it we're going to have messy boundary points like this. It would also allow us to support sub-second bucket spans more naturally than the scaling approach previously envisaged. However, the time to do it is probably once we've upgraded to C++20 when the standard library will have support for parsing durations from text (as well as support for parsing timestamps - see #1459).

[edsavage]

LGTM