[sparse] add sparsity, add wanda sparsifier to ao

[jcaip]

• adds sparsity folder to torchao
• add in wanda sparsifier

I decided to call the folder sparsity because all of the other constructs are called sparsity, it's only torch.ao.pruning (the namespace) that is called pruning, because zafar decided to migrate it. I'm not sure if it's worth the effort to move everything to be called pruning, and given almost everything is called sparsity it seems like the right choice. I do not feel strongly about this though and could be convinced otherwise

[jerryzh168]

I had some discussions with Zafar before, and I remember pruning is the more general name since there is really just structured and unstructured pruning, (don't remember the details though, we could restart the discussion and I may be able to recall). since we are developing something new, should we just use that?

[jcaip]

cc @supriyar @cpuhrsch @jerryzh168

I don't think pruning is the more general name, to me they seem to be used pretty interchangeably. I think sparsity is more widely used then pruning in literature from what I can tell. IMO it doesn't make a difference what name we pick ... that being said if either of you feel differently, I'd be happy to hear thoughts.

[cpuhrsch]

I think this roughly the distinction we had landed on on why pruning is different from sparsity. You could prune, for example, entire MHA heads and just delete them. Sparsity is about storage formats and kernels for efficiently compressing repeat zeros. You can use sparsity to speed up pruned Tensors (sometimes). Sometimes pruning can have advantage even without sparsity.