Recognize regressions as GLMs

I think we could make the existing normal_id_glm implementation available to models written as ordinary normal regressions, such as y ~ normal(alpha + X * beta, sigma), without requiring users to rewrite their models.

I have a compiler PR on my fork that recognizes data-matrix predictors and sums of data-vector predictors. It uses existing Stan Math functions and currently runs only under --Oexperimental. Unsupported forms keep their original calls.

An earlier benchmark reduced whole-run instructions by 56% on one regression. I haven't rerun performance measurements on the revised branch. The current compiler and runtime checks pass, including normalized and proportional densities, gradients, and output names. The GLM changes floating-point association, so I don't expect bit-identical draws.

Would this be useful to bring upstream as an experimental optimization? If there's a better place in the compiler to recognize these forms, or a preferred scope for an initial PR, I'd be happy to adjust it.

Review the compiler PR on the linked fork and the existing normal_id_glm implementation, focusing on its --Oexperimental behavior and supported regression forms. Run the cited compiler and runtime checks, then establish the preferred upstream scope and acceptance criteria for recognizing these models without changing unsupported calls.