Fix GPU build for ptanh_float.

Eigen/src/Core/MathFunctions.h

diff --git a/Eigen/src/Core/MathFunctions.h b/Eigen/src/Core/MathFunctions.h
index aba1d50..a0e9d1a 100644
--- a/Eigen/src/Core/MathFunctions.h
+++ b/Eigen/src/Core/MathFunctions.h
@@ -980,7 +980,7 @@
 template <typename T>
 EIGEN_DEVICE_FUNC bool isinf_impl(const std::complex<T>& x);
 template <typename T>
-T ptanh_float(const T& a_x);
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS T ptanh_float(const T& a_x);
 
 /****************************************************************************
  * Implementation of sign                                                 *

Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h

diff --git a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
index 626185c..b44cb78 100644
--- a/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
+++ b/Eigen/src/Core/arch/Default/GenericPacketMathFunctions.h
@@ -939,7 +939,7 @@
     This implementation works on both scalars and packets.
 */
 template <typename T>
-T ptanh_float(const T& a_x) {
+EIGEN_DEFINE_FUNCTION_ALLOWING_MULTIPLE_DEFINITIONS T ptanh_float(const T& a_x) {
   // Clamp the inputs to the range [-c, c]
 #ifdef EIGEN_VECTORIZE_FMA
   const T plus_clamp = pset1<T>(7.99881172180175781f);