Minor codegen tweaks for ISCA

[gussmith23]

• [X] Figure out which operators are still having code generated for them
  • List:
  • transpose
  • pad
  • flatten
  • access windows
  • squeeze
  • reshape
• [ ] Figure out which operators we can generate less code for
  • transpose This one may be unavoidable. I don't know where it's mostly being done, though. It might get folded into a hardware op. It seems to be happening a lot at the very beginning, to transpose the weights for convolution. that's because we transpose them in our implementation of conv2d. That may not actually be right for how our hardware implements conv2d!
  • pad Can Scott pad in hardware? Can we pad beforehand? Can we pad more efficiently?
  • flatten We don't need to generate code for this!
  • access windows I think Scott can handle the im2col-ing in hardware!
  • squeeze We don't need to generate code for this!
  • reshape We also don't need to generate code for this, just like flatten!
• [ ] Generate less code for operators
  • [ ] transpose
  • [ ] pad
  • [x] flatten
  • [ ] access windows
  • [x] squeeze
  • [x] reshape

[gussmith23]

Some of these "problems" still exist -- they're not exactly problems, but just things we need to implement optimizations for. (For example, we can pre-slice, pre-pad, pre-transpose the data; we can do pointer manipulation to implement slicing, if the sliced data is all contiguous; etc.)

See #91 #92 #93