Teach ManifestArray to handle xarray's reindex/alignment indexer at the
array-protocol layer, so plain xr.reindex, xr.align, and xr.concat work over
virtual arrays and stay lazy — missing positions become null-path manifest
entries that read back as the array's fill_value, with no materialization and
no dtype promotion.

xarray's Variable._getitem_with_mask gathers with an integer indexer (-1 for
missing) then masks with where(). The hooks that make this work lazily:

- manifests/indexing.py: an integer-array indexer is routed to a chunk-grid
  remap (chunk_map_from_indexer -> ManifestArray._reindex_axis) instead of being
  rejected as fancy indexing. Each chunk-block must be all-missing (-> null
  chunk) or a chunk-aligned run (-> real chunk), else NotImplementedError; no
  chunk splitting.
- manifests/reindex.py (new): chunk_map_from_indexer partitions the indexer by
  chunk and asserts the all-or-nothing condition.
- array_api.py: np.where returns the gathered array unchanged when the requested
  fill mask matches its null chunks (the manifest's own fill_value governs);
  general masking still raises. np.result_type treats a lone ManifestArray's
  dtype as authoritative, so int arrays keep their dtype + declared sentinel fill
  rather than being promoted to float+NaN (also fixes a latent
  generator-consumption bug in the previous implementation).
- array.py: __array_function__ now admits ndarray among the argument types
  (np.where's boolean condition arrives as one); adds ManifestArray._reindex_axis.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

A 2D+ outer-join reindexes more than one dimension in a single call, and
xarray expresses that as a broadcast (vectorized) indexer: one integer array
per axis, shaped to the array's rank with the real labels on its own axis and
length 1 everywhere else (e.g. (N,1) and (1,M)). _apply_reindex assumed a 1D
indexer per axis, so it misread the array's rank — raising "axis 0 (chunk
size 1)" or, on newer numpy, TypeError from int() on a non-scalar block.

Add _collapse_outer_indexer: since reindex indexing is orthogonal, each
broadcast component reshapes losslessly to its 1D per-axis indexer. A genuine
pointwise/vectorized indexer (non-axis dim != 1) is rejected with a clear
error. This makes the real ITS_LIVE x+y mosaic work via native reindex/concat.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>