[XLA] Don't return 1 for a random floating point in range [0, 1).

The old implementation of this routine was converting a u32 I to a random f32 F
in range by doing

  uint32 I;
  float F = static_cast<float>(I) / static_cast<float>(2^32);

But this will return 1 if that's the closest float to I / 2^32, and we want a
float in range [0, 1), not [0, 1].

This is significant particularly for normally-distributed random numbers,
because in that case we end up computing erfinv(1) = inf and causing infinities
to appear in our normal distribution.

Instead, we manually construct a floating-point value.

Also fix the interpreter in the same way and add a comment to
HloInstruction::CreateRng.
PiperOrigin-RevId: 233874065