Add `_precompile` and `_precompile_method_instance` intrinsics
The purpose of this is to be able to intercept precompile statements
(currently in codegen and eventually also in inference/inlining), so
that the compiler can respond to these as a "compilation obligation".
For example, consider something like `setindex!(Dict{Any,Int}, ...)`:
```
function setindex!(d::Dict{Any,Int}, obj::Any, idx::Int)
...
# make sure we have `hash!(obj)` ready for us later
Core._precompile(Tuple{typeof(hash!), typeof(obj)})
if (capacity_low)
rehash_all_elements!(d) # this ends up calling `hash!(::Any)`
end
...
return obj
end
```
There is an inherently dynamic dispatch in this code, where we iterate
over all the objects that were _already_ inserted. However, each
particular insertion generally knows what object it is inserting.
The `Core._precompile` statement allows the compiler to use this
type-information at each insertion site and guarantee that we've
compiled sufficient code.
At the dynamic dispatch call-site, `@assert_precompiled hash!(...)`
is used, but we know that the code is available (even in `juliac`)
thanks to the `precompile(...)` statements.
