Use effects tracking to enable better vectorization.

[oscardssmith]

Now that we have the ability to reliably determine that functions are pure, we should try to make our broadcast/map functionality take advantage of this to the extent possible. I'm not 100% sure what needs to be done here, to make sure we generate good code, but it would be nice to get faster performance automatically (at least for the simple stuff) without needing LoopVectorization,

[vchuravy]

I think a relatively low hanging fruit would be to encode effects in LLVM attributes/metadata and then using them in JuliaLICM to move pure math calls out of loops.

Or if someone feels adventurous to pick up #36832 and make it use effects.

Moving function calls out of loops should generally unlock more vectorization potential.