Add a datasets-only mode to `policyengine bundle install` (inverse of --no-datasets)

[anth-volk]

Motivation

policyengine bundle install currently couples two responsibilities: installing the bundle's package scaffold and downloading certified datasets (+ writing the bundle receipt). The CLI offers --no-datasets (packages only) but no inverse — there is no way to obtain just the certified datasets and receipt for a bundle without also having the bundle installer manage packages.

In PolicyEngine/policyengine-api-v2 we are moving the simulation service's Modal images to canonical uv-lockfile installs (PolicyEngine/policyengine-sim-api#609): uv sync --frozen against a committed uv.lock should own all Python packages in the image, so that image environments are byte-identical to the CI-tested environment and can only change through a relock. The bundle's model-package pins are already mirrored into that project's pyproject.toml/uv.lock by automation, so uv can install them — but the bundle install step must still run for the datasets and receipt, and today that drags its package installer along with it. As a bridge we currently run bundle install --venv /.uv/.venv so both installers at least share one environment.

Request

A datasets-only mode, e.g.:

policyengine bundle install <version> --no-packages --data-dir /opt/policyengine/data --yes
# or a dedicated subcommand:
policyengine bundle fetch-data <version> --data-dir /opt/policyengine/data --yes

Behavior: download/verify the bundle's certified datasets for the selected countries into --data-dir and write the bundle receipt, without installing or modifying any Python packages. Symmetric with the existing --no-datasets.

With that available, consumers who manage packages via uv (or any other resolver) can keep the bundle as the source of truth for certified data while their lockfile is the sole source of truth for packages.

[anth-volk]

Additional data point from the integration work: the package half of bundle install shells out to <target python> -m pip install …, which assumes pip exists in the target environment — uv-created venvs don't ship pip, so bundle install --venv <uv venv> fails with No module named pip out of the box (workaround: add pip to the uv dependency group). A datasets-only mode sidesteps this class entirely; alternatively the package installer could fall back to ensurepip or use uv when available.