preserve sign of NaN in conversions to and from BigFloat

[JeffBezanson]

Looking at #5290 led me to finding various issues with the sign bit of NaN, which, let's be honest, is basically what most issues boil down to.

I don't know if this really matters, but our other floating point conversions preserve the sign of NaN so I figure these should too. Currently we get e.g.:

julia> signbit(big(NaN))  # ok
false

julia> signbit(Float64(big(NaN)))  # not ok
true

julia> copysign(1.0, big(NaN))  # not ok
-1.0

julia> signbit(big(-NaN))  # not ok
false

Maybe this is a bug in MPFR; not sure.

[StefanKarpinski]

led me to finding various issues with the sign bit of NaN, which, let's be honest, is basically what most issues boil down to.

🤣

It does seem like MPFR should perhaps handle this better; fixing it ourselves is probably the right approach and maybe filing an upstream bug report.