Fix dpnp.tensor.expm1 handling of complex(+-0, 0) (#2926)

This PR proposes to fix issue #2878 where
`dpnp.tensor.expm1(complex(-0.0, 0.0))` returned `-0.0 + 0.0j` on CPU
instead of `0.0 + 0.0j` as required by the Array API specification

The fix adds an explicit special-case branch for `complex(±0, 0)` in
`expm1.hpp` so CPU and GPU behavior are now consistent and compliant
with the specification.

Also `test_expm1_special_cases` test was extended to cover
`complex(-0.0, 0.0)`

Author: vlad-perevezentsev

CHANGELOG.md

@@ -32,6 +32,7 @@ This release is compatible with NumPy 2.4.5.
 * Fixed missing `libtensor` headers in the installed `dpnp` package [#2915](https://github.com/IntelPython/dpnp/pull/2915)
 * Fixed boolean mask indexing to raise `IndexError` when mask dimensions don't match the indexed array dimensions, aligning with NumPy behavior. Previously, incompatible boolean masks silently returned incorrect results instead of raising an error [#2929](https://github.com/IntelPython/dpnp/pull/2929)
 * Fixed a bug in `astype` where casting floating point types to unsigned integral types could cause an intermediate signed integral type to overflow, leading to incorrect results [#2930](https://github.com/IntelPython/dpnp/pull/2930)
+* Fixed incorrect `dpnp.tensor.expm1` result for `complex(±0, 0)` special case on CPU to match the Python Array API specification [#2926](https://github.com/IntelPython/dpnp/pull/2926)
 
 ### Security
 

dpnp/tensor/libtensor/include/kernels/elementwise_functions/expm1.hpp

@@ -121,6 +121,10 @@ struct Expm1Functor
                 }
             }
 
+            if (x == realT(0) && y == realT(0)) {
+                return resT{realT(0), y};
+            }
+
             // x, y finite numbers
             const realT cosY_val = sycl::cos(y);
             const realT sinY_val = (y == 0) ? y : sycl::sin(y);

dpnp/tests/tensor/elementwise/test_expm1.py

@@ -147,6 +147,7 @@ def test_expm1_special_cases():
     num_finite = 1.0
     vals = [
         complex(0.0, 0.0),
+        complex(-0.0, 0.0),
         complex(num_finite, dpt.inf),
         complex(num_finite, dpt.nan),
         complex(dpt.inf, 0.0),
@@ -165,6 +166,7 @@ def test_expm1_special_cases():
     c_nan = complex(np.nan, np.nan)
     res = np.asarray(
         [
+            complex(0.0, 0.0),
             complex(0.0, 0.0),
             c_nan,
             c_nan,
@@ -184,4 +186,10 @@ def test_expm1_special_cases():
 
     tol = dpt.finfo(X.dtype).resolution
     with np.errstate(invalid="ignore"):
-        assert_allclose(dpt.asnumpy(dpt.expm1(X)), res, atol=tol, rtol=tol)
+        Y = dpt.asnumpy(dpt.expm1(X))
+        assert_allclose(Y, res, atol=tol, rtol=tol)
+
+    # assert_allclose treats +0 == -0
+    # verify sign bits for zero real parts
+    for i in (0, 1):
+        assert not np.signbit(Y[i].real)