Allow collect()'ing when all processing is complete

[unode]

Scenario:

1. 12 samples are being processed using the parallel machinery lock1() and collect().
2. 10 samples complete and 2 fail.
3. The 2 failing samples are considered bad and are excluded from the sample file.

At this point re-running ngless has no effect since all work is complete however the merged output from collect() was never generated.

collect() can also fail to occur in rare cases where the last two samples finish almost simultaneously or filesystem lag prevents the last two processes from seeing all samples as complete.

[unode]

In order to keep compatibility with the current behavior (no action when finished), I'm wondering if this should be implemented through a --only-collect command-line option.

Effectively we have to skip all actions (preprocess, map, fastq, paired, ...) except collect but, we still need to have a sample name for collect to act upon.

[luispedro]

Over the long term, I would prefer an approach where, whenever ngless runs¸ it will create any missing outputs. The whole lock1/collect business is a bit of a hack now. This is probably for NGLess 2, though.