Add GitHub action to check model running using `cmdstanr`

[athowes]

We could have a GitHub action to check that our Stan code works with the latest version of cmdstanr.

See #65.

This is what the action: https://github.com/epinowcast/epinowcast/blob/main/.github/workflows/check-cmdstan.yaml is doing inside of epinowcast.

To make this change we would need to copy over this, change the parts that are epinowcast specific. For sure this part:

https://github.com/epinowcast/epinowcast/blob/886b45cb4bc5f338fa53d22e83d25335c55b1a4a/.github/workflows/check-cmdstan.yaml#L60-L77

      - name: Compile model and check syntax
        run: |
          model <- epinowcast::enw_model()
          # If the model is not syntactically correct above will fail
          # however it may be correct enougth to compile but still contain
          # soft depreciated syntax and so we check the syntax again below 
          # and test the output.
          message <- capture.output(
            model$check_syntax(pedantic = FALSE),
            type = "message"
          )
          # We can't use TRUE here as pendatic check return lots of false
          # positives related to our use of functions.
          stopifnot(
            length(message) != 0 &&
            all(message == "Stan program is syntactically correct")
          )
        shell: Rscript {0}

Do we need to agree on which functions will be tested they can run? What about user input Stan functionality? Basically how many possible variations should we be checking?

Also I don't think we have a function to just compile, so we'd be also running. I imagine this would slow things down. Perhaps there is a way we can use fn to just compile and not run?

[seabbs]

We use make_stancode to get the standcode and pass that to cmdstanr to compile here I think.

I think we just want to check that our inserted stan makes a valid stan model using the base brms settings and trust brms to otherwise produce valid stan code.

[athowes]

I wonder how this is different to having a unit test which checks the code compiles, which we already have?

(Thinking of having a go at closing this issue now.)

[seabbs]

It can not fail but still have depreciation warnings essentially. Its proved useful to pull compiling out of unit tests as it makes clear what is failing.

[athowes]

By pull out do you mean to remove the compliation unit test? Or you think both serve useful purpose?

[athowes]

This is the code that needs to be replaced with something from epidist:

model <- epinowcast::enw_model()
message <- capture.output(model$check_syntax(pedantic = FALSE), type = "message")
# We can't use TRUE here as pendatic check return lots of false positives related to our use of functions.
stopifnot(length(message) != 0 && all(message == "Stan program is syntactically correct"))

It would look something like:

stancode <- epidist::epidist(data = ?, fn = brms::make_stancode)
mod <- cmdstan_model(stan_file = write_stan_file(stancode), compile = FALSE)
expect_true(mod$check_syntax())

The data here isn't doing anything, so I don't know what we can do to give it something so that it works (without actually generating plausible data). So long as it passes as_latent_individual I suppose.

[seabbs]

Yes - you can see the approach we took in the delays paper for generating stan code as a guide.