Thanks, but I don't think so. On one hand, feature spaces in machine learning are usually
not conceptual vector spaces (e.g. in a tall/short classification, it doesn't make sense to add two "tall"s together or multiply a "tall" by a scalar). In general, features have different data types: some booleans, some integers, some real numbers. We only lump them together, cast them into a homogeneous data type and store them in a multidimensional array solely for interpolation purpose.
On the other hand, since you come from a math background, I suppose you know the concept of "free vector space" (which is usually introduced along with the universal mapping property of tensor product). As said in a previous comment, technically every object can be viewed as a vector, and hence a tensor. A duck is a tensor, a bird that quacks like a duck is a tensor, a Java SimpleBeanFactoryAwareAspectInstanceFactory is a tensor, a Neo Armstrong Cyclone Jet Armstrong Cannon is a tensor, a red-black tree is a tensor. But we don't call them tensors, because normally we don't do anything tensorial to them.
Then why shall we treat multidimensional arrays differently? Apart from the mere fact that a tensor in physics/mathematics can be stored numerically as a multidimensional array, my impression is that (please correct me if I'm wrong) the coinage of the term in machine learning libraries is a complete disregard for the original physical or mathematical concept. Is "multidimensional array" a name so hard to understand that we need a new name? Does the new name in any way enhance one's understanding of machine learning that the old name cannot? I wonder.