[tensorflow/tensorflow] tf.math.reciprocal_no_nan returns inconsistent results for identical complex inf+0j inputs

### ROOT CAUSE The issue arises because `tf.math.reciprocal_no_nan` does not consistently handle complex infinity inputs. The function uses `std::isnan` to check for NaN, but this does not detect infinity. As a result, for complex infinity inputs, the function falls back to computing the reciprocal, which returns 0 for real infinity but may produce inconsistent results for complex infinity due to floating-point precision or system-specific behavior. ### CODE FIX ```cpp // In the kernel implementation for reciprocal_no_nan, add a condition to check for infinity if (std::isnan(x)) { *y = x; } else if (std::isinf(x) || std::abs(x) == std::numeric_limits::infinity()) { *y = 0; } else { *y = recip(x); } ``` This ensures that any input (real or complex) that is either NaN or infinity is handled consistently by returning the input for NaN and 0 for infinity.