While working on implementing a new inliner for ZJIT, I keep running into an issue where the wrong iseq pointer is being used. Since `JITState#iseq` and `FrameState#iseq` end up being the same value in the absence of inlining, I don't think we've been all that rigorous about using the "correct" one in a given context. I've been migrating to `FrameState#iseq` in my inliner work and adapting on each rebase. But, having looked at this code for a little while, I think we can remove `JITState#iseq`.

This PR switches to using `FrameState#iseq` in cases where we're working with the code associated with a specific frame (e.g., local-table layout, side-exit resumption target) rather than the overall compilation unit. Other cases are cleaned up by either storing the iseq on the instruction or computing the value that required the iseq and storing that in the instruction.

Without inlining, this change set is really just a clean-up of what we have today. With inlining, however, these changes become semantically meaningful. Making the change now removes the decision point of when to use `JITState#iseq` vs `FrameState#iseq` and simplifies keeping the inliner work current with _master_.


* ZJIT: Use `FrameState#iseq` for variable-access codegen

* ZJIT: Use `FrameState#iseq` for `IsMethodCfunc` codegen

* ZJIT: Precompute LEP level for defined?(yield) at HIR construction

Rather than use `JITState#iseq` to compute the LEP level during codegen, we can compute it once and store it at HIR construction.

* ZJIT: Use `FrameState#iseq` for local spills and side-exit reconstruction

Previously, we were using `JITState#iseq`, but these call sites are more about the frame state rather than the outer compilation unit.

* ZJIT: Remove the unused iseq field from JITState