[let-def]

This PR implements two changes.

First a bug fix in Cmm_helpers.preallocate_block which dropped block descriptions.
In particular, this broke cyclic constants:

 --- Cyclic List ---
-[1; <cycle>]
+(1, <cycle>)
 
 --- Prefixed cyclic list ---
-[0; 1; <cycle>]
+(0, (1, <cycle>))
 
 --- Indirect cycle ---
-[1; 2; 3; <cycle>]
+[1; 2<malformed list>]

The second fix is an attempt at making optimization passes "block_desc"-aware, but I am not sure whether this is a good idea or not.

@lthls: If you have some time to spare, I would be grateful if you could take a look ;)

[lthls]

The Cmm_helpers part is obviously right (it should have been part of my Flambda PR, but I missed it).
The flambda changes are more subtle, but they both correspond to things that I thought about when writing my PR so I can give some context.

The code in Simple_value_approx was intentionally dropping block descriptions when encountering blocks with the same contents but different descriptions. This was meant to keep optimisations that share identical blocks (so that adding introspection would not degrade performance). But I think it was actually misguided: this code isn't involved in constant sharing, only when merging approximations coming from different branches (it should have been called join, not meet, but that's another story).
I don't know if it matters in practice: I think that we can only get matching tags and sizes but mismatched descriptions when GADTs are involved, and in that case it's probably fine to lose the approximation. So I think we can keep your patch (at least when introspection is enabled).

The code in Simplify_primitives looks more dangerous, as getting it wrong would be a soundness issue. We use structurally_different to prove physical difference, so if two blocks only differ in their descriptions I chose to err on the side of caution and assume that they might actually get shared (if the descriptions get dropped, for instance). I don't think there is a significant benefit to being more precise here, so I would advise to take the less risky road and leave the code as I wrote it.

[let-def]

Thanks for your quick feedback. I have dropped the structurally_different part. The negative names should have tipped me off that it was a conservative approximation.

I will make a new release asap including your flambda changes.