Performance regression due to absence of lto

[wds15]

Description

The freshly merged static matrix refactor things do slow down stan programs significantly whenever lto is not turned on.

So as of now in the static matrix world we absolutely need the lto optimisation for good performance. Thus, our makefiles should ideally switch on lto whenever the compiler is capable enough of handling it. The user should not be required to add -flto to the make/local to get good performance

Example

See #2007 (comment)

where it is demonstrated that turning on lto gives back the old performance numbers.

Expected Output

Results should be produced as fast as in the pre static matrices.

Current Version:

v3.3.0

[SteveBronder]

So yes I'm working on this right now. I decided to toss the idea of using flto for clang-6 or 7 (it's very hard to tell which version you are on). So as long as the user is using gcc >= 5.0 and clang > 7 we should have everything plus a few extra wizzes and woos we can turn on

It's taking me a minute because a few compiler settings + flto can actually lead to a slow-down(!) / wrong values on the performance tests. My guess is that too many optimizations + flto is causing too much inlining and some other algebra optims can cause slightly different numerics

[SteveBronder]

@wds15 can you throw the model and data up into a gist? If not feel free to email it to me

[rok-cesnovar]

This will be fixed with an upstream PR, so reopening this until that happens.