Address fehiepsi review: remove get_log_density_fn, inline via initialize_model

[juanitorduz]

Addresses fehiepsi's two latest comments on pyro-ppl#2203. Scoped so the diff shows only the review-feedback changes (6 files).

Base branch

Base is `external-samplers-master-merge` = `external-samplers` + a routine merge of `origin/master` (which moved the repo to the ty-only toolchain: mypy removed, `ty.toml` deleted, ty config in `pyproject.toml`; the lone `pyproject.toml` conflict was resolved by keeping master's `[tool.ty.*]` and adding `numpyro/infer/util.py` to `[tool.ty.src].include`). Keeping that merge out of this diff is intentional, so reviewers see just the feedback changes here.

Addressing the review

Both comments asked to drop a thin abstraction; done via full removal + inline:

• Removed `get_log_density_fn`, its `LogDensityInfo` return type, and the `PositionDict` alias.
• The log density is now built inline by negating `model_info.potential_fn` (`initialize_model` already returns a single-position `potential_fn`/`postprocess_fn` with `dynamic_args=False`), per "we can add 2 lines of code in the notebook instead."
• Updated `numpyro/infer/init.py`, `docs/source/utilities.rst`, the notebook prose, and the tests. Kept the `constrain_fn` (`batch_ndims`) work and the end-to-end `MCMCKernel` coverage.
• `numpyro/infer/util.py` is now type-checked by ty; this surfaced 3 pre-existing diagnostics in `find_valid_initial_params` (`w.message` typed `Warning | str`), fixed with a one-line `assert isinstance(w.message, Warning)` narrowing.

Verification

• `make lint` green (ruff check, ruff format, header check, `ty check`).
• `prek run --all-files` all hooks pass.
• `pytest test/infer/test_external_helpers.py test/infer/test_infer_util.py` → 61 passed.
• `notebooks/source/other_samplers.ipynb` fully re-executed (Pathfinder + MCLMC + flowMC), 0 errors, outputs refreshed.

🤖 Generated with Claude Code