Make sure dependencies are up to date in travis builds

[ivirshup]

Travis builds are currently breaking due to incompatible versions of numpy being present, so no compatible version gets installed. This should fix that.

[ivirshup]

Not completely sure if this is doing what I intended. I added the -U so dependencies would be upgraded, but numpy still isn't being upgraded as shown by these warnings:

umap-learn 0.4.6 has requirement numpy>=1.17, but you'll have numpy 1.15.4 which is incompatible.
scvi 0.6.6 has requirement numpy>=1.16.2, but you'll have numpy 1.15.4 which is incompatible.

Not sure why this is happening. I'd prefer if we didn't have to manually specify the dependencies of our dependencies. Any ideas @flying-sheep?

---

Updating pip doesn't seem to do anything (maybe it has to do with "editable mode"?)

---

An easy fix is just to add a version requirement on numpy, but I really feel like dependency resolution should be dealing with that.

[ivirshup]

Okay, this is weird. My current theory is that it's a bug in pip. It looks like if multiple packages depend on a single dependency, pip doesn't necessarily check if all requirements are satisfied. It just checks if one packages version requirements are. It seems non-deterministic which package is used to check.

---

I was working on an example, but then I just found this: pypa/pip#8218

There's an unstable feature for this (--unstable-feature=resolver), which does work, but I think adding a requirement for numpy is a bit more justifiable.

[flying-sheep]

I’m really happy that they’re finally adding a resolver.

Without officially having a resolver, it technically wasn’t a bug that incompatible stuff got installed, but instead just a missing (if very vital) feature.