Raise AlignmentError when alignment would reorder an already-aligned index (closes #10714)

[Kayvan-Zahiri]

Closes #10714.

The bug

align is documented to raise AlignmentError when objects carry conflicting
indexes, but the consistency check in _get_dim_pos_indexers only runs for
indexes that need reindexing:

if obj_idx is not None and self.reindex[key]:

_need_reindex returns False precisely when the matching indexes are already
equal, so such an index pins its dimension in place while producing no
re-indexer of its own. Nothing is ever registered for that dimension, so a second
index sharing it has nothing to conflict with and silently wins.

Using @abaarsma's MVCE, x is identical in both objects and xb is not:

ds1 = Dataset(coords={"x": [1, 2, 3], "xb": ("x", [4, 5, 6])}).set_xindex("xb")
ds2 = Dataset(coords={"x": [1, 2, 3], "xb": ("x", [4, 6, 5])}).set_xindex("xb")

align(ds1, ds2)   # no error

It is not only a missing error. ds2.xb is [4, 6, 5] going in and [4, 5, 6]
coming out: the conflicting index is silently rewritten and the mismatched index
is used to combine the data. align(ds1, ds3), where both indexes need
reindexing, raises correctly today, so the difference is purely whether the
second index happened to need reindexing.

The fix

Record the dimensions pinned by non-reindexed indexes, then reject any re-indexer
that would reorder one of them.

The identity check matters and is not redundant. _need_reindex can return True
for reasons other than index inequality, notably differing unindexed dimension
sizes, and in that case reindex_like returns an identity indexer that reorders
nothing. Treating "needs reindexing" as "conflicts" would break those cases, so
the guard is specifically "would reorder", not "would reindex".

dim_pos_indexers is unchanged, so the actual reindexing behaviour is untouched.
The only difference is that a previously silent case now raises.

Verification

• The MVCE now raises; align(ds1, ds3) still raises; align(ds1, ds1) and
  align(ds1) are unaffected.
• Reordering a dimension is still legal when nothing else pins it, which the new
  test asserts explicitly rather than leaving to inference.
• test_indexes.py, test_concat.py, test_merge.py: 300 passed, 2 skipped.
• Every file containing an align test: 1279 passed, 433 skipped, 14 xfailed,
  4 xpassed. Two failures in test_chunk_by_season_resampler are an ImportError
  for cftime and reproduce identically on an unmodified checkout, so they are
  environmental rather than caused by this change.
• The new test fails with DID NOT RAISE on main and passes with the fix.

Credit

The diagnosis is @abaarsma's, including pinpointing the exact branch in
_get_dim_pos_indexers, and @dcherian confirmed it. Nobody had opened a PR, so I
have written it up with a regression test and a whatsnew entry.

AI Disclosure

• This PR contains AI-generated content.

  • I have tested any AI-generated content in my PR.

  • I take responsibility for any AI-generated content in my PR.

    Tools: Claude (Claude Code), used to diagnose the root cause, draft the fix and the
    regression test, and draft this description.