Now it works like this:
- test- branches and pull requests are built with conda but not uploaded.
- tags are built with conda and uploaded with `tags` label.
- cron builds are built with conda and uploaded with `nightly` label.

Nightly builds are assignd a version number equal to branch.date
for version branches or 2.99.date for master branch, e.g.

2.5.190603
2.99.190603

In this way they always supersede the corresponding releases.

See #412

@gtribello I have to revert this since it fails on travis.
Can you please fix it and add it again on v2.4 (then merge, as you did)?

Thanks!

This reverts commit ce4bbf4e.

- Allow Plumed objects to be explicitly finalized
(can be used to make sure files are closed correctly)
- Allow Plumed objects to be used in context managers

Check that METAD + TARGET + GRID_SPARSE works correctly.

Solves a bug with METAD + TARGET + GRID_SPARSE
See also #467

cc: @gtribello

It looks like the change was implemented separately in 2.4 and 2.5

should be not important. In particular, when invsigma was <10-12
sigma was set as equal to invsigma. Now it is set to zero,
which I think is the expected behavior.

When conda build fails, the travis job should return an error.