Still todo:

• Update [ tests so that all the [ tests are evaluated in 3 contexts:
  with R values, with eager tensors, and with graph mode tensors in a tf_function().
• Update the warning and errors thrown when a 0 is encountered.

This is slightly more complicated than I initially estimated.
The complication is a slice like x[6:1] in R. There is seemingly no way to translate that to python, without turning the slice end into a None. That’s easy enough with R values or eager tensors, but is difficult to do in a tf.cond() with placeholder tensors because tf.cond requires that all branches return the same types.

You can get past the tf.cond() by constructing a potentially empty tf.experimental.Optional with a varient dtype, or using a tf.raw_ops.OptionalNone(), but then Tensor.__getitem__() won’t accept that as input.

There are two ways around, both with tradeoffs.
One is to wrap the call to Tensor.getitem in a series of tf.cond() or tf.switch(), for each combination of axes doing something like if end == 0: x[start::step] else x[start:end:step]. This doesn't seem right, and it seems like it would introduce significant performance implications by making a barrier for async TF operations.

Another alternative it to manually construct the end_mask bit mask as a tensor, and then calling tf.strided_slice() directly, bypassing Tensor.__getitem__(). This is a substantial refactoring and will have to wait for the time being.

Follow up: tf.strided_slice() does not accept tensors for begin_mask, end_mask, or ellipsis_mask arguments, only ints. This means that these arguments must be fully known at tracing time, and cannot be dependent on graph values. Consequently, building up a tf.strided_slice() call in the R [.tensorflow.tensor method is not a viable approach for us here.