Support for phantom bindings

[lthls]

This PR restores support for generating phantom bindings.

I wrote it by enabling phantom lets, and fixing errors as I find them. Given the number of bugs I had to fix, it is likely that some remain in some less-tested corner cases.

Here is a summary of the changes:

• The Name_mode lattice is now a straight line. The typing env requires all the variables in it to have a name mode at least In_types, so I made Phantom greater than In_types instead of not comparable. It would also be possible to prevent the typing env from handling variables in Phantom mode, and that should make the environments smaller, but I think it should be a separate PR.
• The logic in Expr_builder to decide whether to keep a binding, delete it or phantomise it has been moved to Simplify_let_expr.rebuild_let. This makes decisions more consistent (now even bindings removed by mutable unboxing get phantomised if needed)
• Slot offset computations now ignore slots not in normal mode, not only in projections. Phantom closures also get Deleted code IDs to avoid keeping old code around.
• No CSE equations are created for phantom bindings

It is possible that this PR still contains changes that were made redundant by a more general fix. Do not hesitate to point these out during review.

[chambart]

Ok this looks sound.
It's hard to tell whether anything is missing out of the changes that would lose any bindings. But the changes themselves seems ok.
There are a few places where It's quite hard to know if the name modes are actually the right one, but I don't know how to verify that without too much effort. I suppose that trying it in the wild is quite safe, as if anything is wrong on that part, this is going to fail quite loudly with all the information we need.