[WIP] Set up CNN accelerator codegen skeleton + counting utilities

[slyubomirsky]

It took a little bit of hacking to get attributes working, but I have an efficientnet import example and a codegen skeleton for the CNN accelerator. Once we have the ASM worked out, it should be pretty easy to do all the plumbing.

Right now the runtime will only print things; it won't actually do anything because I don't know what to target.

There's a small function in ilacnn.py that rewrites any conv2d with padding into conv2d(pad(data), weight) so that padding will be handled entirely by our codegen. In theory, the runtime can do that part implicitly, but doing this tiny rewrite saves us a ton of debugging.

The script for running the EfficientNet trial is a little wacky. Let me know if there's anything I can better implement about it or how to improve it. Note that you will need mxnet installed for the import to work (I don't know if the dockerfile has it), though otherwise everything is handled seamlessly by TVM's importer.

Also includes some simple utilities for counting the number ops, overloads, etc

Please review and comment @AD1024 @LeeOHzzZ @Bo-Yuan-Huang @ttambe

[Bo-Yuan-Huang]

Can the new updates be merged to a branch under uwsample/3la-tvm or can we have access to the private fork slyubomirsky:hlscnn? So that we can try out the code. Thanks!

[slyubomirsky]

Oh, i thought you would already have access. You are welcome to fork my branch (or to push to this PR). I will give you all appropriate permissions