Support libdevice.frexp

[isuruf]

Describe the issue

Currently I use the following, but it is suboptimal.

@triton.jit
def frexp(x):
    y = libdevice.ilogb(x) + 1
    exponent = tl.where(x == 0, 0, y)
    mantissa = tl.where(x == 0, 0, libdevice.ldexp(x, -y))
    return  mantissa, exponent

There is a libdevice function __nv_frexp https://docs.nvidia.com/cuda/archive/10.2/libdevice-users-guide/__nv_frexp.html but it has a pointer as an argument as there are two results. However the functions in libdevice.py all have a single output, so I'm not sure how to implement this triton.

Environment details

Triton: 3.1.0