Revert to old default scale constraint in autoguides

[fritzo]

This reverts autoguide scale constraints from #2767 to the old exp-transformed versions until @vitkl can perform experiments confirming improved performance. If experiments succeed, we can simply revert this PR. Note this PR preserves the ability to easily perform those experiments, via

old_autonormal = AutoNormal(model)
old_autonormal.scale_constraint = constraints.positive

new_autonormal = AutoNormal(model)
new_autonormal.scale_constraint = constraints.softplus_positive

See also pyro-ppl/numpyro#941

[fehiepsi]

I like this decision. :)

[vitkl]

I like the decision - but did any issues with this come up?

[fritzo]

did any issues with this come up?

yes, some inference tests started failing in #2767 and pyro-ppl/numpyro#941. To get those tests to pass in #2767 I tweaked some learning parameters. those failures are not strong evidence for either constraint, but in the absence of strong evidence, we should preserve existing behavior.

[fritzo]

@fehiepsi could you merge this?