qemu-ppc (user) incorrectly converts float(nan)->double(non-nan)

Noticed on qemu-3.1.0 on GHC test suite where float32 comparisons didn't work on NaNs.
Here is the minimal reproducer:

```c
// cat a.c
#include <stdio.h>
#include <math.h>
#include <stdint.h>

int main() {
    volatile float f1 = NAN;
    volatile float f2 = NAN;
    printf ("f1 (%e, %#x) >= f2 (%e, %#x): %s\n",
        f1, *(volatile uint32_t*)&f1,
        f2, *(volatile uint32_t*)&f2,
        (f1 >= f2) ? "True"
                   : "False");
    volatile double d = f1;
    printf ("d (%e, %#llx)\n",
        d, *(volatile uint64_t*)&d);
}
```

```
# incorrect execution:
$ powerpc-unknown-linux-gnu-gcc -O2 a.c -o a -static && qemu-ppc ./a 
f1 (5.104236e+38, 0x7fc00000) >= f2 (5.104236e+38, 0x7fc00000): True
d (5.104236e+38, 0x47f8000000000000)

# correct execution
$ scp a timberdoodle.ppc64.dev.gentoo.org:~/;  ssh timberdoodle.ppc64.dev.gentoo.org ./a
f1 (nan, 0x7fc00000) >= f2 (nan, 0x7fc00000): False
d (nan, 0x7ff8000000000000)
```

Note: qemu-ppc handled float32 extension as it was not a NaN (exp=111..1111) but a normalized number.



The bug is in the same area as https://bugs.launchpad.net/qemu/+bug/1821444 but in another branch of 'uint64_t helper_todouble(uint32_t arg=0x1)'.

We also hit this regression when testing CompCert for e5500 with qemu 3.1.0.

The minimal example
> #include <stdio.h>
> #include <math.h>
>  
> union C { float f; unsigned long l; };
> int main (void) {
>     union C c;
>     c.f = NAN;
>     printf("Float: %f\n  Hex: 0x%x\n", c.f, c.l);
>     printf("The above float is %s NAN\n", c.f == NAN ? "==" : "!=");                                                      
>     return 0;
> }

does not exhibit the error with GCC 6.3.0, since that statically optimizes the c.f == NAN to always be false. But CompCert doesn't do this optimization and thus triggers the regression in the float compare.

With qemu 3.0.0 the example outputs (whether built with compcert or gcc):
> Float: nan
>   Hex: 0x7fc00000
> The above float is != NAN

The example works well with qemu 3.0.0 and we stumbled over this with qemu 3.1.0; both build from the released source tarballs. My native system is a x86_64 GNU/Linux (openSUSE Tumbleweed with Kernel 4.20.13-1-default).

Output with qemu 3.1.0 and built with GCC:
> $ ./qemu-ppc64abi32-3.1.0 bug25310.gcc.elf 
> Float: 510423550381407695195061911147652317184.000000
>   Hex: 0x7fc00000
> The above float is != NAN

qemu 3.1.0 and example built with CompCert:
> $ ./qemu-ppc64abi32-3.1.0 bug25310.ccomp.elf 
> Float: 510423550381407695195061911147652317184.000000
>   Hex: 0x7fc00000
> The above float is == NAN

I attached example binaries (statically built):
> $ ../../compcert_internal_ppc/bin/ccomp --version
> The CompCert C verified compiler, Release: 18.10i, Build: 4503984, Tag: auto/2019/02/11/1924
> $ ../../compcert_internal_ppc/bin/ccomp -target e5500-linux -g -static bug25310.c -o bug25310.ccomp.elf 
> 
> $ ../../compcert_internal_ppc/gcc/bin/powerpc-unknown-linux-gnu-gcc --version
> powerpc-unknown-linux-gnu-gcc (crosstool-NG crosstool-ng-1.23.0) 6.3.0
> $ ../../compcert_internal_ppc/gcc/bin/powerpc-unknown-linux-gnu-gcc bug25310.c -g -static -O0 -o bug25310.gcc.elf -mcpu=e5500