Extend adelie to support different penalties

[JamesYang007]

Other softwares seem to support a lot more penalties such as sparse group lasso (l1 + l2), MCP, SCAD. Think about how to abstract out penalty and constraints. This is tricky because a constraint class currently dictates how to perform the block-update, but it hard-codes the penalty to be elastic net.

Nice to haves:

• [ ] The constraint classes should not be so tied to the penalty type.

API:

ad.grpnet(
    X=X,
    glm=ad.glm.gaussian(y),
    penalty=ad.penalty.elastic_net(alpha, penalty_factor),
    constraints=constraints,
)

• [ ] We should be able to detect given a penalty and a constraint whether there exists a block-update function given these two.