Fix #207

[mhauru]

Attempting to fix #207. The current version fixes the MWE, but I've failed to make it into a test (the MWE stops crashing once you wrap it in a testset). Hoping to run the Turing tests against this though.

[github-actions]

Libtask.jl documentation for PR #209 is available at:
https://TuringLang.github.io/Libtask.jl/previews/PR209/

[mhauru]

Local testing indicates that the particle MCMC tests in Turing should pass with this (at least most do), and the performance gains are kept. Still need to figure out how to do the test case for this in Libtask's suite.

[Technici4n]

the MWE stops crashing once you wrap it in a testset

You could maybe try to look at the IR you get for the function and see what the type of the phi node is? Testsets are supposed to behave like global statements so I'm a little surprised.

[mhauru]

I found a variant of the MWE that fails in a testset too. @Technici4n, would you be happy to review? If this makes sense to you we could release straight-away, assuming CI comes out fine.

[Technici4n]

LGTM!

[mhauru]

The integration tests are still running, but I did the crucial parts locally and so far they seem fine on CI too, and I need to run off now until Monday, so I'm just gonna risk it. It went so well with #205.