Add StableHLO complex sqrt to stablehlo-complex-math-expander pass

[pearu]

As in the title.

The existing implementation of sqrt on complex inputs uses polar form of complex sqrt which is inaccurate/incorrect on about 28 % of uniformly distributed samples over all complex plane. The JAX complex sqrt accuracy statistics is as follows:

test_unary[sqrt-jax-cuda-complex64-default] maximal ULP difference: 2792619172
ULP difference == 0: 297588
ULP difference == 1: 1106945
ULP difference == 2: 78921
ULP difference == 3: 5924
ULP difference == 4: 2938
ULP difference == 5: 2231
ULP difference == 6: 1855
ULP difference == 7: 1648
ULP difference == 8: 1449
ULP difference == 9: 1263
ULP difference == 10: 1089
ULP difference >= 11: 598949

test_unary[sqrt-jax-cpu-complex64-default] maximal ULP difference: 2760370645
ULP difference == 0: 699582
ULP difference == 1: 759750
ULP difference == 2: 58127
ULP difference == 3: 3237
ULP difference == 4: 1665
ULP difference == 5: 1185
ULP difference == 6: 1021
ULP difference == 7: 871
ULP difference == 8: 804
ULP difference == 9: 653
ULP difference == 10: 622
ULP difference >= 11: 573283

This PR provides an algorithm for complex sqrt that is accurate up to 3/6 ULP difference error on complex samples. The corresponding JAX complex sqrt accuracy statistics is as follows:

test_unary[sqrt-jax-cuda-complex64-default] maximal ULP difference: 5
ULP difference == 0: 1060571
ULP difference == 1: 1008268
ULP difference == 2: 31136
ULP difference == 3: 686
ULP difference == 4: 129
ULP difference == 5: 10

test_unary[sqrt-jax-cpu-complex64-default] maximal ULP difference: 2
ULP difference == 0: 1348868
ULP difference == 1: 751504
ULP difference == 2: 428

It is interesting to note that although the same algorithm is used for both CUDA and CPU platforms, then the expected maximal ULP difference is 4 (obtained from applying the algorithm to numpy arrays). Hence

• the accuracy of complex sqrt on CPU is better than expected
• the accuracy of complex sqrt on CUDA is worse than expected because CUDA sqrt produces slightly different results from std sqrt on float inputs.

[pearu]

@GleasonK , please review.