Privatize build_vrt; route to_geotiff's VRT path through it

[brendancol]

Summary

Closes #2974.

• Drops build_vrt from the public GeoTIFF API and renames the wrapper to _build_vrt (out of __all__ and the package docstring).
• Routes to_geotiff's .vrt write path (_write_vrt_tiled) through _build_vrt, so it's the single internal entry point for VRT-index emission. That path used to call _vrt.write_vrt directly, which left the public build_vrt with no internal caller at all.
• Updates the docs (GeoTIFF reference, safe-IO guide, release-gate table, internal notes), the README feature matrix, and the user-guide notebook's VRT section to write a mosaic through to_geotiff('.vrt') instead of the removed helper.
• Updates tests to import _build_vrt; the public-surface assertions now expect only open_geotiff and to_geotiff.

build_vrt is advanced-tier, not stable, so it's allowed to change between releases. Anyone who called xrspatial.geotiff.build_vrt either writes to a .vrt path through to_geotiff or imports _build_vrt.

Backend coverage

No behavior change. VRT writing already ran on the eager and dask paths; this is an API-surface refactor. Round-trip checked: to_geotiff(da_chunked, 'mosaic.vrt') tiles the array and open_geotiff reads it back matching the original.

Test plan

• Public-surface assertions flipped to {open_geotiff, to_geotiff}, plus a new test that build_vrt is not importable while _build_vrt is
• Every test file that imported the public name now uses _build_vrt
• VRT round-trip through to_geotiff('.vrt') passes on numpy and dask
• flake8 + isort clean on changed files
• User-guide notebook VRT cell re-run; output regenerated from a real execution

[brendancol]

PR Review: Privatize build_vrt; route to_geotiff's VRT path through it

Blockers

None.

Suggestions

None blocking.

Nits

• _build_vrt's docstring opening now reads [internal-only], but the five per-parameter tier tags below it still say [advanced] (_writers/vrt.py:58-90). Minor internal inconsistency now that the function is no longer a public entry point; worth normalizing to [internal-only].

What looks good

• The routing change is behavior-preserving. _build_vrt(vrt_path, tile_paths, relative=True, nodata=nodata) resolves to write_vrt(vrt_path, tile_paths, relative=True, crs_wkt=None, nodata=nodata), the same call the path made before. No deprecation warning fires (path positional, crs unset), and the local import sidesteps a circular dependency.
• Public-surface assertions are flipped correctly, and the new test_build_vrt_is_not_public covers both directions: the public name is gone and _build_vrt is still importable.
• The pre-existing local helper _build_vrt(tmp_path) in test_finalization.py would have collided with the imported _build_vrt after the rename; it was renamed to _make_one_source_vrt, and the imported call inside it was left intact.
• The notebook's VRT output cell was regenerated from a real run rather than hand-written, so the printed tile sizes and round-trip result are accurate.

Checklist

• No algorithm change (API-surface refactor)
• Backends unaffected; VRT write still runs eager + dask
• No NaN / edge-case concerns
• Routing change verified behavior-preserving
• No premature materialization or extra copies
• Benchmark not applicable
• README feature matrix updated
• Docstrings present (one tier-tag nit above)

[brendancol]

Follow-up review

The one nit from the first pass is fixed in d04a252: the five per-parameter tier tags in _build_vrt's docstring now read [internal-only], matching the function summary. test_polish.py::TestC5WriteVrtKwargs still passes (the docstring keeps relative / crs_wkt / nodata).

No new findings. The change is docstring-only; no code paths moved.