Fix overflow for hypot in Float16

[simonbyrne]

Fixes #38311, includes tests from #38272.

[KristofferC]

Bump, here's another one from PkgEval:

julia> hypot(Float16(0.004406) - Float16(0.01092)im)
NaN16

[JeffBezanson]

Bump. Looks like this has some test failures?

[JeffBezanson]

Bump.

[stevengj]

Wouldn't we be better off doing the Float16 calculations as:

_hypot(x::Float16, y::Float16) = Float16(_hypot(Float32(x), Float32(y)))
_hypot(x::ComplexF16, y::ComplexF16) = Float16(_hypot(ComplexF32(x), ComplexF32(y)))

Though I guess we can do that in addition to this PR, since there might be other types (e.g. quaternions) defined in packages whose absolute value gives a Float16.

[simonbyrne]

I think that is the best way forward. I don't think it is possible to handle Float16 without at least some special casing, simply because the subnormal range is so wide compared to the exponent range.