Fix code age relation so that unsimplified code is not sent to Expr_builder

[mshinwell]

At present, Lambda_to_flambda is not computing full free name information on terms, which means that Expr_builder will not operate correctly if passed unsimplified code for a new "let symbol" binding. (For example it won't see uses of code IDs in function bodies, leading to code bindings being erroneously deleted.)

The reason this was happening turned out to be a non-trivial code age relation, resulting from significant inlining, looking like:

    code_id1  [earliest, output from Lambda_to_flambda]
       |
    code_id2  [first piece of simplified code]
      /  \
     /    \
code_id3  code_id4

The check that the code age relation remained linear was being too conservative: it was saying that the relation branched after code_id1, which corresponded to unsimplified code, and thus such unsimplified code had new "let symbol" bindings created for it (not marked as Deleted). This in turn meant that earlier "let symbol" bindings whose bound code IDs were used in the body of the code associated with code_id1 were deleted, since not all of the free names in that code were visible to the earlier bindings. It should in fact suffice to say that the relation only branched after code_id2, which was the first piece of simplified code; that is what this PR does.

This will be superceded by ocaml-flambda/ocaml#556 which eliminates this check on the shape of the code age relation in favour of a different analysis.

However we need to have a discussion about the free name information computed by Lambda_to_flambda. If we're not going to compute full free name information there, we may need to change the types so that "free names unknown" can be represented, to make this more robust. cc @lthls @chambart