[XLA:CPU] Use vectorized implementation for all calls to log/exp/tanh.

Previously, we used llvm.{log,exp,tanh}.f32 when we failed to vectorize and for
the tail of vectorized loops.  The latter was particularly problematic because
the LLVM intrinsic doesn't have the same numeric properties as our vectorized
implementation, and so we'd get different numerics for different output
elements!

PiperOrigin-RevId: 231869262