Explicit setting of the default backend

[phinate]

It may be desirable in some cases (e.g. #882) to have one consistent backend for all cases, which would enable the propagation of gradients throughout pyhf model construction, for instance. Backends will hit a wall when trying to take the gradient with respect to something that involves pure numpy.

An example implementation of this could be as simple as adding the argument set_default=False to set_backend, followed by something like

def set_backend(..., set_default=False):
    ...
    backend = tensorlib
    if set_default:
        global default_backend
        default_backend = tensorlib
    ...
    tensorlib._setup()
    if set_default:
        default_backend._setup()

It seems like this would be fine at first glance, since the backend checks are done in the lines before this. I managed to get the logpdf of hepdata_like with no issues, but I was unsure what would be a more robust test for changing the default backend. Ideally, one would want to run the whole test suite I guess, to make sure this doesn't secretly change some functionality.

Hope this is useful!

[kratsg]

The default backend can be set w/o changing the rest of the code. It'll help with the diff'able case for sure. My first guess is you should be able to do

import pyhf
pyhf.default_backend = pyhf.tensor.tensorflow_backend(precision='64b')

at the top of your code? I'm not 100% sure one wants an updated API for this, but...