Improve ci runs

[nickva]

A few improvements to the test suite to help it run better in parallel.

The main improvement is to separate the per-app makefile targets early, during templating phase, so they never share a common app structure during runtime. That lets us avoid reconfiguring the data directory at runtime, which effectively is an almost full restart internally since we're tearing down all the couch servers. So we're saving an extra app shutdown + restart cycle per app with the new scheme and improving the isolation.

The new scheme uses the tmp/$app namespace for each app. After the tests run any unsuccessful tmp/$app directories will be left behind so they can be inspected. This is especially useful for looking at the failed couch.log entries. There is a separate catlogs target to concatenate failed app couch.log files though maybe keeping them separately may be even easier to inspect. Since if app a failed, we'd just rather look at app a's logs and not app c's logs.

Since setup_eunit rebar template now takes an app argument, opted to remove some older eunit targets keeping only the new ones supporting parallel execution. Otherwise, would have to add a special case statement in the plugin for no-app vs one app cases.

Previously, make eunit apps=a,b,c did not run apps a, b and c as separate goals in parallel execution. They were passed in as the a,b,c goal. To fix that parse the apps into individual space separated goals, so they can
be sped up by parallel execution as well.

To keep the apps as isolated from each other as possible make build_eunit_config also create separate etc folders, use separate BUILDDIR macros and each apps only uses its own fixtures. Also use 0-valued ports theeunit template to avoid any chance of port collision and resulting flakiness.

Rename SUBDIRS to EUNIT_SUBDIRS since they used for eunit tests currently only. Also, exclude docs and fauxton apps from the target list, since they are not proper erlang apps and don't have eunit tests.

Newer Erlangs have file:del_dir_r/1 so use that in cleanup_dirs to save a few lines of code.

There are a few other separate commits to update application with a few timeouts I noticed when running with higher concurrency. Another is make the replicator use its own fixture for attachments and not dip into couch app's fixtures folder since those could be setup up or torn down separately and we would rather not add intra-target dependencies during parallel runs.

Let's try running with failfast=false for a bit in Jenkins since it might make it easier to debug failures. Once things settle, or if we see a lot of backlog from failing (doomed run) we can switch back.