perf: short-circuit COO.reshape when -1 resolves to self.shape

COO.reshape returns self when self.shape equals the requested shape,
but only checks before resolving any -1 in the target. sparse.tensordot
passes shapes like (-1, N) even for 2D x 2D matmul that doesn't actually
change shape, so the short-circuit never fires and a full reshape runs
(linear_loc + coord rebuild + new COO allocation).

Moving the -1 resolution before the equality check avoids that work for
callers that pass a -1 factorization of the current shape. Behavior is
a strict subset ("fewer copies"): any reshape that already returned
self before still does, and reshape((-1, ...)) that resolves to the
current shape now also returns self, matching the documented contract.

Measured ~16% speedup on a warm conservative-regrid loop (xarray-regrid
ConservativeRegridder.regrid) whose tensordot call sits on the hot path;
bit-identical output. All 6050 existing numba-backend tests pass.

Author: thodson-usgs

sparse/numba_backend/_coo/core.py

@@ -1070,11 +1070,11 @@ def reshape(self, shape, order="C"):
         if order not in {"C", None}:
             raise NotImplementedError("The `order` parameter is not supported")
 
-        if self.shape == shape:
-            return self
         if any(d == -1 for d in shape):
             extra = int(self.size / np.prod([d for d in shape if d != -1]))
             shape = tuple([d if d != -1 else extra for d in shape])
+        if self.shape == shape:
+            return self
 
         if self.size != reduce(operator.mul, shape, 1):
             raise ValueError(f"cannot reshape array of size {self.size} into shape {shape}")