BUG: expose broadcast_shapes in the Dask namespace

[Ra1nhy]

Fixes #439.

Export the existing dask.array.core.broadcast_shapes through the wrapped namespace. This adds no new broadcasting algorithm and preserves Dask's unknown-dimension handling rather than routing shapes through NumPy. The new Dask-specific regression verifies that known and unknown shapes can be combined without computing an array or changing its chunks.

Remove the name from the missing-export list and remove the five corresponding standard-test xfails after verifying them with --max-examples 200. This is only a missing metadata-function export, not a change to SciPy's broader Dask support policy or distributed execution.

Validation on macOS, CPython 3.12.14, Dask 2026.8.0, NumPy 2.5.4, based on 482d5f9:

• Both new no-compute cases fail with AttributeError before the export and pass afterward.
• Local project tests with the available dependencies: 272 passed, 461 skipped, compared with 270 passed, 461 skipped on untouched upstream. The same existing NumPy overflow warning occurs on both.
• Five focused Array API tests (names, signature, compatible/incompatible shapes, no arguments): 5 failed before, 5 passed after, with the property tests configured for 200 examples. Test source: array-api-tests@6c0b59f, specification stubs: array-api@5f847a3. They ran from a fixed-source export; a test-runner namespace bootstrap loads the unchanged 2025.12 stubs without importing unexported historical stub versions. No test function or 2025.12 stub was modified.
• Ruff 0.17.1 check passed for the three Python files.

Other array backends, other Python/Dask versions, the complete Array API suite, type checking and documentation builds were not run locally.