Make fitted tests backend-neutral (#311, #328)

[beckyfisher]

Addresses #311 and #328.

What

Twelve fitted tests no longer pass open_progress = FALSE through bnec() or
bnec_hurdle(). The argument is specific to the rstan backend. Omitting it
lets the same tests run when brms is configured to use cmdstanr.

The branch is based on dev after #331 merged the #328 fixture reuse. That
change builds five Stan programs where the affected files previously built
fourteen. This pull request completes the combined test-infrastructure batch
with #311. It also corrects stale descriptions of #328's withdrawn worker-count
policy and prevents a manually dispatched timing run from truncating a
fractional worker count.

Why it matters

The test suite previously failed when a contributor had
options(brms.backend = "cmdstanr") in the session or profile. The failure was
reported as a model that did not fit, although the model and data were valid.
The retained fixture reuse removes nine repeated compilations without changing
an assertion.

Evidence

On R 4.6.1 with brms 2.23.0, cmdstanr 0.9.0 and CmdStan 2.39.0, a revised
nec4param test fit compiled and sampled through the cmdstanr backend. Both
chains completed and bnec() returned a bayesnecfit. All eight edited test
files parse, no open_progress call remains under tests/, and
R CMD build --no-build-vignettes succeeds.

The workflow parses as YAML. Direct tests of its worker-input rule accept
positive integral counts and reject fractional, non-finite, out-of-range and
non-positive values.

The fixture reuse merged in #331 was measured directly: the affected files
compiled five Stan programs rather than fourteen with assertion counts unchanged
at 71 and 36. Its four-platform R CMD check matrix passed. This pull request's
matrix tests the combined state.

Implementation detail

open_progress defaults to interactive() for the rstan sampling path, which
is already false in non-interactive test runs. The cmdstanr path does not
accept the argument. Deleting it therefore changes no model equation, family,
prior, seed, chain count, iteration count or assertion.

#328's addendum withdrew the proposed CI worker-count change after paired runs
could not distinguish its effect from runner variance. This branch does not
restore that setting. It retains only #331's deterministic fixture reuse and
the workflow instrumentation that already merged to dev. A blank dispatch
input leaves TESTTHAT_CPUS unset, and a supplied value must represent a
positive integer before conversion.

No scientific judgement was required for either part.

[beckyfisher]

The pull request body now includes the corrected #328 worker-count description and validation of manually supplied worker counts. The latter rejects fractional, non-finite, out-of-range and non-positive values before integer conversion.

[beckyfisher]

Independent review completed after three cycles. No defect was found in the
test changes or their scientific behaviour.

The review identified and corrected these maintenance problems:

• The workflow said that a blank dispatch input used a runner-computed worker
  count, although it leaves TESTTHAT_CPUS unset. The description now states
  that testthat uses its default of two workers.
• Manual dispatch silently truncated a fractional worker count. Validation now
  rejects fractional, non-finite, out-of-range and non-positive values before
  integer conversion.
• The work queue retained Ten of the suite's twenty-nine Stan compilations rebuild a program already built #328's withdrawn worker-count plan and stale pull
  request status. It now records Build each repeated test fixture once: fourteen Stan compilations to five (#328) #331's merged fixture reuse, Make fitted tests backend-neutral (#311, #328) #349's scope, the
  withdrawn policy, and the current open pull requests.

The reviewer accepted the corrections and found no remaining actionable issue
at 99ad0115.