A4: Implement reshape fast path for _unstack_once

[FBumann]

Finding

Investigation under #3 isolated where the 48% of multiindex_stack workload time inside _unstack_once actually goes. Result:

variable shape: (20, 40000) dtype: float64
multiindex levels: [200, 200] is_monotonic: True is_unique: True
new_shape: (20, 200, 200) is_missing: False

Time per step (over 20 iterations):
 transpose only : 0.002ms
 allocate template only : 0.001ms
 allocate + scatter write : 0.920ms ← all of it
 (alternative) reshape : 0.000ms ← essentially free
 full _unstack_once : 0.912ms

Reshape == scatter result? True

100% of the time is in data[(..., *indexer)] = reordered — numpy advanced-indexing scatter. But for the common case (lex-sorted MultiIndex, no missing combinations — i.e. a normal .stack(...).unstack(...) round trip) the codes form a row-major sequence and a plain reshape produces the same array.

Proposed fix

Two-line guard in xarray/core/variable.py::_unstack_once:

if not sparse and not is_missing_values and index.is_monotonic_increasing:
 # Complete lex-sorted Cartesian product — codes are row-major,
 # so reshape produces the same result as the scatter write.
 data = reordered.data.reshape(new_shape)
else:
 data = create_template(self.data, shape=new_shape, dtype=dtype)
 data[(..., *indexer)] = reordered.data

Existing scatter path stays as the fallback for sparse output, missing combinations, or a user-constructed non-lex-sorted MultiIndex.

Expected impact

• _unstack_once: ~900× speedup on stack→unstack round trips (the common case)
• multiindex_stack workload: ~48% of total time eliminated
• Zero behavior change on the fallback paths

Acceptance

• Existing variable + multiindex tests pass
• New test covers both fast path (lex-sorted, complete) and fallback (missing or non-lex-sorted) producing identical results
• asv benchmark for stack/unstack shows the speedup

Closes #3.

---

[This is Claude Code on behalf of Felix Bumann]