Fix transform_tensor for non-default allocator float_vec

fplus::transform with rvalue relies on type deduction that fails when
float_vec uses Eigen::aligned_allocator. Replace with an explicit copy
and in-place std::transform, which achieves the same effect: the copy
uses memcpy rather than zero-initialization, and the plain iterator
allows the compiler to vectorize the loop.

Author: Dobiasd

include/fdeep/tensor.hpp

@@ -190,7 +190,9 @@ namespace internal {
     template <typename F>
     tensor transform_tensor(F f, const tensor& m)
     {
-        return tensor(m.shape(), fplus::transform(f, float_vec(*m.as_vector())));
+        float_vec result(*m.as_vector());
+        std::transform(result.begin(), result.end(), result.begin(), f);
+        return tensor(m.shape(), std::move(result));
     }
 
     inline std::vector<tensor> tensor_to_depth_slices(const tensor& m)