[agourakis82]

Complex asin discards the imaginary component on the real-axis branch outside [-1, 1] and returns a positive real component for negative arguments. This also gives incorrect results in acos, which is defined through asin.

For example, asin(2 + 0i) returns approximately (1.5708, 0) instead of (1.5708, 1.31696), while asin(-4 + 0i) returns a positive half-pi instead of a negative one. Restore the signed real component and the imaginary magnitude using real acosh, with the branch side selected by the imaginary zero. Preserve the input imaginary zero inside [-1, 1]. Clamp the real acosh argument to its domain in lanes where that branch is discarded.

Add mixed-lane float/double regressions for asin and acos: both imaginary-zero signs, signed real zero, interior points, endpoints, adjacent values outside the endpoints, and off-axis controls. Comparisons use scalar C++ complex functions, with exact zero/sign assertions where applicable.

Validation (C++23):

• ARM64: identical new tests fail before the patch (276 assertions); the final combined trigonometric/hyperbolic test units pass (6 cases, 1056 assertions).
• ASan/UBSan: the final combined units pass.
• Native GCC 13.3 SSE2 and AVX2: baseline failures and final-candidate passes confirmed for both units.
• The unchanged original Complex asin, acos, asinh, acosh, sqrt, log1p, tan, tanh return very different results from std:: #1432 reproducer now matches its scalar references for the reported real-axis asin/acos cases and its imaginary-axis asinh examples.
• A separate local composition check with Select the principal branch of complex acosh in both half-planes #1436 passes 8 cases / 1632 assertions, including the original acosh(2 - 0i) example. This PR is independently based on master; the acosh branch selection remains in Select the principal branch of complex acosh in both half-planes #1436.
• Formatting and whitespace checks pass.

Addresses part of #1432. Large off-axis cancellation/overflow, nonfinite inputs, and complete architecture coverage remain outside this patch. The new signed-zero assertions are disabled under fast-math; no fast-math signed-zero guarantee or full-suite pass is claimed. Floating-point exception freedom is not claimed: the clamp preserves the baseline FE_INVALID state for a focused interior-axis witness, including a pre-existing float FE_INVALID flag.