Refine effects based on optimizer-derived information (#50805)
The optimizer may be able to derive information that is not available to
inference. For example, it may SROA a mutable value to derive additional
constant information. Additionally, some effects, like :consistent are
path-dependent and should ideally be scanned once all optimizations are
done. Now, there is a bit of a complication that we have generally so
far taken the position that the optimizer may do non-IPO-safe
optimizations, although in practice we never actually implemented any.
This was a sensible choice, because we weren't really doing anything
with the post-optimized IR other than feeding it into codegen anyway.
However, with irinterp and this change, there's now two consumers of
IPO-safely optimized IR. I do still think we may at some point want to
run passes that allow IPO-unsafe optimizations, but we can always add
them at the end of the pipeline.
With these changes, the effect analysis is a lot more precise. For
example, we can now derive :consistent for these functions:
```
function f1(b)
if Base.inferencebarrier(b)
error()
end
return b
end
function f3(x)
@fastmath sqrt(x)
return x
end
```
and we can derive `:nothrow` for this function:
```
function f2()
if Ref(false)[]
error()
end
return true
end
```
