Typeclasses and dimensions-0.3.2.0

[stites]

See #39 for description. Closes #18.

The private-to-public switch did something to my fork and blocked me from updating #39, so I'm closing it in favor of this PR.

[austinvhuang]

Let's get this merged @stites, can you submit the PR to a staging branch so we can resolve fixes without breaking the master branch?

[stites]

Yep. So I just rebased this and need to review some changes. Issues I saw in the rebase:

• TensorRaw started porting code to do operations on RawTensorLong. This is unnecessary as the generic functions can handle all tensor types.
• nn got deleted in master???
• examples are still broken due to compile errors. Also hasktorch-examples:exe:ad changed significantly, so reverted back to master's version. See next comment

[stites]

everything builds and stack test works, sans standard_gamma functions, which are new in ATen. I commented out these functions for now so that they would build on my machine which has an older .so, but I think it would also make a good first issue for someone to reintroduce this code -- @omaclaren, maybe?

@austinvhuang tell me what you think about the PR -- I think it's ready at this point, even though FloatMath has been commented out (it seems to be a c/p of DoubleMath, and I'd prefer to port DoubleMath to a generic version, like how Static.Generic is really a refactored Static.Double). The code as-is shouldn't be considered complete, there is still a lot of reorganizing that I think needs to happen before a release, but merging this moves us back to smaller PRs.

[austinvhuang]

Merged. Great work @stites! The aten function update should be easily patched, and as I mentioned in slack, it's probably time for an upstream refresh since they've been adding some other probability density related functions to TensorMath.

[stites]

👍 cool beans! I'm also getting a better sense of what the typeclass-based interface would look like (it's not super crazy) for the higher-level Dynamic and Static modules.