Today we don't support a reduction that results in a single scalar value, as the current infrastructure would see this as resulting in a zero-dimensional tensor.

We should first support zero-dimensional tensors, meaning be able to create them and pipe them through the code generator so it's implicitly recognized as a single scalar value and we can generate code for them.

This may include:

• Modifying the TensorView class
• Making sure arith.[c,h] can take 0-dim tensors as inputs and operate with scalar values (resulting in 0-dim tensor outputs)
• Modifying the computeAt logic
  • Make sure zero-dim doesn't result in any asserts
  • Make sure we can still call computeAt(..., 0) on a 0 dim tensor
• Checking tensor replay doesn't error if it gets a 0 dim tensor unless computeAt is somehow > 0
• Checking index computations are correct for consumer/producer global/local
• Checking we can re-broadcast the value with and without blockBroadcast
• Creating tests