Flaky: NativeLinAlg.dot matrix-matrix ComplexFloat returns wrong result on Linux

### Multik version

0.3.0

### Module

multik-openblas

### Platform

JVM

### Operating System

Linux

### Description

`NativeLinAlg.dot` for matrix-matrix multiplication with `ComplexFloat` produces incorrect results intermittently on Linux (Ubuntu 22.04.5 LTS, gcc-12 12.3.0, JDK 21 Temurin). The test multiplies small integer ComplexFloat matrices and asserts with `assertComplexFloatingNDArray` using epsilon `1e-4f`. Failed in 2 out of 4 CI runs.

### Reproducer

```kotlin
val matrix1 = mk.ndarray(
mk[mk[ComplexFloat(1, 2), ComplexFloat(3, 4)],
mk[ComplexFloat(2, 3), ComplexFloat(4, 5)],
mk[ComplexFloat(3, 4), ComplexFloat(5, 6)]]
)
val matrix2 = mk.ndarray(
mk[mk[ComplexFloat(9, 8), ComplexFloat(7, 6), ComplexFloat(5, 4)],
mk[ComplexFloat(8, 7), ComplexFloat(6, 5), ComplexFloat(4, 3)]]
)

val actual = NativeLinAlg.dot(matrix1, matrix2)
assertComplexFloatingNDArray(expected, actual, 1e-4f) // Fails intermittently
```

### Additional information

```text
Flaky — failed in 2/4 CI runs.
AssertionFailedError at NativeLinAlgTest.kt:156.
CI runs: 22511673326, 22511861936.
Environment: Ubuntu 22.04.5 LTS, gcc-12 (12.3.0), g++-12, JDK 21 Temurin, x86_64.
```

Start with the failing assertion at NativeLinAlgTest.kt:156 and the matrix-matrix ComplexFloat reproducer in the issue. Inspect the NativeLinAlg implementation in the multik-openblas module and compare behavior across the listed Linux CI environment. Done means the reproduced multiplication is stable and the test passes repeatedly with the stated epsilon.

Written by the indexing model from the issue text.