The non-signalling versions of VSX scalar convert to shorter/longer
precision insns doesn't silence SNaNs in the hardware. To better match
this behavior, use the non-arithmatic conversion of helper_todouble
instead of float32_to_float64. A test is added to prevent future
regressions.

Signed-off-by: Matheus Ferst <matheus.ferst@eldorado.org.br>
Message-Id: <20211228120310.1957990-1-matheus.ferst@eldorado.org.br>
Signed-off-by: Cédric Le Goater <clg@kaod.org>