make tests have their own environment

[tpapp]

Introduces a separate test environment to reproduce the problem in #94 (which is unrelated to that PR), then adds a fix by rearranging the branches and testing for infinites early.

I still don't understand why the previous setup was not using the latest ForwardDiff.jl, but this one does.

[oscardssmith]

Can you revert the test Project changes? IMO the [extras] way of managing versions works better (and TestEnv.jl can be used to activate the test environment), since the separate Project.toml approach is easy to have issues where someone updates one but not the other, leading to issues.

[tpapp]

I don't agree, I think that with the [extras] you can easily run into mystery issues like not catching this and using an ancient ForwardDiff.jl. But I am open to arguments and you can convince me to revert that part if necessary.

[oscardssmith]

I don't especially care. I think this should merge.

[devmotion]

I think the correct fix for mysterious test environments would be to add compat entries for test environments. This is what a currently missing Aqua test would require as well.

[oscardssmith]

yeah compats for test env is a good idea in general.

[tpapp]

I added QA with Aqua and JET. I am merging this as is, so that #94 can proceed, if this causes issues we can go back to having test packages in the main Project.toml.

[devmotion]

I think this change should be benchmarked. IIRC the performance of logsumexp was very sensitive to even minor changes of the implementation. To avoid the multiplication one might want to use

Suggested change

	if isinf(real_x) && isinf(real_xmax) && (real_x * real_xmax) > 0
	if isinf(real_x) && isinf(real_xmax) && sign(real_x) == sign(real_xmax)

[devmotion]

IMO test dependencies should also be specified with compat entries and this file should be updated with CompatHelper.