Add support for HRF convolution

[tyarkoni]

For a variety of reasons, it makes sense to move HRF convolution out of fitlins and into pybids. This will make it much easier to construct "final" design matrices we can just hand off to estimators like nistats. We can achieve this very easily by bundling @bthirion's hemodynamic_models module from nistats in pybids.

@bthirion, any objections? We could also add nistats as a dependency, but I don't think it makes sense to do that at the moment given that this is the only thing we need from nistats, and I'd like to avoid adding sklearn and nilearn as dependencies if we can help it.

[bthirion]

I understand your concern: reducing the dependencies ! On the other hand, I'd surprised you don't use Nilearn in your data processing (if you don't you probably should ;-))

[tyarkoni]

pybids doesn't do any data processing; right now it never even reads an image (well, except to determine the number of volumes). I'd like to keep it that way if I can help it... :)

[effigies]

Would it make sense to just make the design matrix bits a small, self-contained package that pybids and nistats could each depend on? It should have an order of magnitude less churn than the rest of nistats, as it's (almost) exactly the same as what was in nipy, before.

[bthirion]

I'd rather not: we realize that package maintenance is a burden, and we prefer to have fewer, well maintained packages rather than a bunch of small ones that nobody will take care of.
For this we reason, we'll actually merge Nistats into Nilearn in the next months (Yes, I've been saying that for months, but now we have somebody actively working on it) !