# Assume Tensors and DataFrames are builtin to the system (i.e. they produce code)
# based on backend selections). 

# A Tensor is a container of a Type. Zero-dimensional tensors exist
#  and are analogous to elements of the Type.
#  You must index the 0-d array to produce the element.

# Elements of the Tensor must be explictly extracted using the syntax a[()]
# where a evaluates to a 0-d Tensor.  This always copies the element into a new 
# object.

# A Tensor has attributes: shape, ndim, dtype, T, kind

# A Tensor kind is a meta-type which is one of the Types:
# General > Gamma > Strided > C-contiguous > F-contiguous
# General > Chunked 

# Based on the kind of Tensor we have private fields that are present or not
# GeneralMeta:
    index_map : a function of ndim arguments and a state Object that produces
                a 0-d tensor. 
# General:
    index_map : (a function of ndim integer arguments that produces a 0-d tensor)
# Gamma adds:
    gamma_map : (a function of ndim integer arguments that produces an integer)
    datapointer : a 1-d array of bytes
# Strided adds:
    strides : a Tuple of integers indicating the number of bytes to jump to get
# C-Contiguous:
    no additional information, the other attributes functions are computed as needed.
# F-Contiguous:
    no additional information, the other attributes are computed as needed.
    
Also builtin to the system are "generalized universal functions"

These operate on