Fix: GCC 9.2 fast-math complex products

GCC 9.2 miscompiles packetized coefficient-based complex products under compiler -ffast-math, which can make a lazy row-major complex product disagree with the evaluated result.

Disable that packet path only for real GCC versions before 9.3 when FAST_MATH is active, leaving newer compilers and non-fast-math builds unchanged.

GCC's likely upstream fix is PR tree-optimization/92420, "Vectorization miscompilation with negative strides", backported to the GCC 9 branch as commit 2d8ea3a0a609 before GCC 9.3.0. That fix is not in GCC 9.2.0, matching the observed failure boundary.

Add a fastmath regression covering row-major complex adjoint products for both lazy auto evaluation and explicit column-major evaluation.

Fixes: #1839 (closed)