Reduce mish error by an alternative without softplus op

[ChinChangYang]

Fix the high numerical error in mish activation #2359.

Algorithm:

e = exp(x)
mish = x / (1 + 2 / (e * (e + 2)))

Evaluation:

In the following experiments, the mean absolute errors are evaluated by the method in #2359 (comment).

Before this change, NE generates high numerical error:

Mean Absolute Errors Across Samples:
  var_17:
    NE:  2.955052
    GPU: 0.000998

With the new algorithm, NE generates low numerical error:

Mean Absolute Errors Across Samples:
  var_17:
    NE:  0.001744
    GPU: 0.001516

A tester reported that the new mish function generates NaN only when x is -Inf in the float16 format.

Performance:

This change has been adopted in KataGo Core ML backend ChinChangYang/KataGo#7. The performance of the KataGo model with the new mish activation (7.15 ms) is similar to the original mish implementation (7.03 ms).

Conclusion:

Overall, the change enhances the accuracy and reliability of the mish activation in Core ML models.