Mark complex exp, sinh, and cosh infinite border regions as inaccurate in lax_test.
https://github.com/openxla/stablehlo/pull/3039 updates StableHLO's
complex math expander lowering of `stablehlo.exponential` to follow C99
Annex G.6.3.1 when the real part is infinite and the imaginary part is
non-finite: `exp(-inf + I*y)` returns `0` and `exp(+inf + I*y)` returns
`+inf + nanj`. Because complex `sinh` and `cosh` lower through
`stablehlo.exponential`, `sinh(±inf ± infj)` evaluates to `±inf + nanj`
and `cosh(±inf ± infj)` evaluates to `+inf + nanj`.
Meanwhile, `mpmath` still evaluates all three operations at
`(±inf ± infj)` to `nan + nanj`, causing `testSuccessOnComplexPlane` for
`exp`, `sinh`, and `cosh` (`complex64` and `complex128`) in
`FunctionAccuracyTest` to fail on CPU and GPU.
Mark `'ninf'`, `'pinf'`, `'ninfj'`, and `'pinfj'` as known inaccurate
regions for `exp` and `cosh`, and expand `'ninf.imag'` and `'pinf.imag'`
to `'ninf'` and `'pinf'` for `sinh` on CPU and CUDA (`'ninfj.real'` and
`'pinfj.real'` are already listed for `sinh`). This remains
backward-compatible with older XLA/StableHLO builds because
`testFailureOnComplexPlane` skips regions that still match `mpmath`.
PiperOrigin-RevId: 1024054142