[mlir][arith] Arith folder for maxnumf/minnumf has wrong NaN behaviour

[oowekyala]

IEEE754-2019's maximumnumf and minimumnumf have NaN as their identity element. The folder was copied from the regular maximumf and minimumf and considers that infinities are the identities.

For instance

%1 = arith.maxnumf %nan, %something : f32

should fold to %something but

%1 = arith.maxnumf %-inf, %something : f32

should not be folded, as if %something is NaN, then %-inf should be returned.