This commit limits the lifetimes of `OptimizationState` and `IRCode`
for a more dataflow clarity. It also avoids duplicated calls of `ir_to_codeinf!`.

Note that external `AbstractInterpreter`s can still extend their
lifetimes to cache additional information, as described by this
newly added documentation of `finish!`:

>     finish!(interp::AbstractInterpreter,
>         opt::OptimizationState, ir::IRCode, caller::InferenceResult)
>
> Runs post-Julia-level optimization process and caches information for later uses:
> - computes "purity" (i.e. side-effect-freeness) of the optimized frame
> - computes inlining cost and cache the inlineability in `opt.src.inlineable`
> - stores the result of optimization in `caller.src`
> * by default, `caller.src` will be an optimized `CodeInfo` object transformed from `ir`
> * in a case when this frame has been proven pure, `ConstAPI` object wrapping the constant
> value will be kept in `caller.src` instead, so that the runtime system will use
> the constant calling convention
>
> !!! note
>     The lifetimes of `opt` and `ir` end by the end of this process.
>     Still external `AbstractInterpreter` can override this method as necessary to cache them.
>     Note that `transform_result_for_cache` should be overloaded also in such cases,
>     otherwise the default `transform_result_for_cache` implmentation will discard any information
>     other than `CodeInfo`, `Vector{UInt8}` or `ConstAPI`.

This commit also adds a new overload `infresult_iterator` so that external
interpreters can tweak the behavior of post processings of `_typeinf`.
Especially, this change is motivated by the need for JET, whose post-optimization
processing needs references of `InferenceState`.