Add the hubverse model output interface

[nickreich]

The reason for the rewrite: hubverse model output goes in, allocation scores
come out.

alloscore_model_out() takes a model_out_tbl of quantile forecasts and the
matching oracle_output and scores them, replacing roughly thirty lines of
per-analysis boilerplate — nest ps/qs, deframe,
add_pdqr_funs(dist = "distfromq"), join truth — that was reproduced in every
downstream script. See utility-eval-papers/R/run-alloscore.R for the version
this displaces.

scores <- alloscore_model_out(
  model_out_tbl = dplyr::filter(hubExamples::forecast_outputs, output_type == "quantile"),
  oracle_output = hubExamples::forecast_oracle_output,
  K = c(500, 1000, 2000),
  target_cols = "location",
  by = c("model_id", "K")
)

The one genuinely new design decision

An allocation problem is a set of targets sharing one budget, and hubverse has
no single column for that. So target_cols names the task ID columns whose
combinations enumerate those targets, and the allocation unit is derived as
model_id plus every remaining task ID column. One problem is solved per
combination.

That mirrors how a hubverse compound_taskid_set works: the columns you name
vary within a group, the rest are held constant. With
target_cols = "location" on the hubExamples forecast data you get 24
allocation problems (3 models x 2 reference dates x 4 horizons), each pooling 2
locations.

This is the part most worth review attention — it is an interpretation of the
hubverse data model, not a mechanical port.

Functions

• as_alloscore_df() — exported, because it is where a mis-specified join or
  target set becomes visible. Returns one row per allocation unit with a
  forecasts list column holding each target's quantiles, its distfromq-built
  cdf and quantile function, and the observed outcome.
• allocate_model_out() — stops after the allocation, for when the question is
  what a forecast implies you should do rather than how good it was.
• alloscore_model_out() — allocates and scores, with summarize/by in the
  shape hubEvals::score_model_out() uses.

Validation

Follows hubEvals: get_task_id_cols() and validate_model_oracle_out()
mirror its helpers, and a single supported output_type is enforced up front.
Only quantile is accepted; sample, cdf, pmf, mean and median error
with an explanation rather than silently doing something wrong.

Rejected inputs covered by tests: several output types at once, a non-quantile
output type, non-numeric quantile levels, oracle_output with unexpected
columns or missing coverage, duplicated forecasts, and target_cols naming
something that is not a task ID.

Tests

418 total, up from 306, in two commits.

The interface adds a test worth looking at,
"alloscore_model_out matches the core API on a single unit": it routes one
allocation problem through both the hubverse entry point and the core
alloscore() and asserts the scores agree, which is what makes this a refactor
of the interface rather than a second implementation.

The equivalence fixtures then prove both agree with the original package.
The original has no test suite, so there is no pre-existing ground truth to
port: data-raw/make_legacy_reference.R runs the original code once over a
fixed set of fixtures and freezes the results under
tests/testthat/testdata/, and test-legacy-equivalence.R checks this package
reproduces them. The original therefore stays out of this package's
dependencies while the numerics stay pinned, so any future change in behaviour
has to be deliberate.

legacy_hub_quantile_scores.csv is the one that matters here: hubverse example
data pushed through the original's hand-rolled pipeline, which
alloscore_model_out() must reproduce. It ships alongside the function rather
than in a later PR, so the interface does not land unproven.

legacy_reference_meta.json records the original's git SHA (0477ae9) and the
R, distfromq and hubExamples versions used. Tolerances are 1e-6 on
allocations and 1e-8 on scores. A mutation check confirms the suite is not
vacuous: widening eps_K from 1% to 5% fails these tests and the ZXH tests,
and restoring it passes all 54.

Verified locally

R CMD check: Status: OK. 418 tests, 0 failures, 0 warnings, 0 skips.
lintr: 0 lints. air format . --check: clean.