gxf_isaac_gems: make MachineEpsilon specializations inline

[wolfv]

epsilon.hpp declares a variable template and two explicit specializations:

template <typename K>
constexpr K MachineEpsilon = K(0);
template <>
constexpr float MachineEpsilon<float> = 5.9604644775390625e-8f;
template <>
constexpr double MachineEpsilon<double> = 1.11022302462515654e-16;

In C++17 constexpr implies inline for variables, but an explicit specialization of a variable template is not a template -- it is an ordinary variable declaration, and the implicit inline does not apply to it. Each translation unit including the header therefore emits its own definition with external linkage, and linking two such objects together fails:

ld: detection3_d_array_message.cpp.o:(.rodata+0x38): multiple definition of
    'nvidia::isaac::MachineEpsilon<double>';
    nitros_detection3_d_array.cpp.o:(.rodata+0x48): first defined here
ld: detection3_d_array_message.cpp.o:(.rodata+0x40): multiple definition of
    'nvidia::isaac::MachineEpsilon<float>'; ...
collect2: error: ld returned 1 exit status

Adding inline gives the specializations vague linkage so the definitions are merged, which is what the primary template already gets.

Hit while building isaac_ros_nitros_detection3_d_array_type from source with GCC 15 and binutils ld. It is latent rather than new: any two translation units in one target that both include epsilon.hpp will collide, so it depends on how the sources happen to be grouped.