Initial refactoring of To_cmm

[mshinwell]

There should be no changes here at all except refactoring. Best read commit by commit. The main things are:

• To_cmm_helper is no longer a dependency of To_cmm_env and To_cmm_result. This means we can share more functions between different source files.
• The primitive simplification code has been moved to To_cmm_primitive.
• The code for dealing with the Flambda_unit.t itself is now in To_cmm and the old To_cmm was moved to To_cmm_expr.
• The recursive binding in To_cmm_expr has been significantly shrunk in size, which I think helps to understand the flow of what's going on, although it does mean that some of the helper functions are now further away. Various functions have been simply inlined out; it was kind of hard to follow what was going on before, and this also reduces the risk of variable names diverging (which they had in some places).
• Some minor changes such as removing unnecessary res parameters in some places.

One of the main aims is to break the code down into smaller pieces so it's easier to check critical invariants, e.g. relating to linear use of the environment.

[Gbury]

Mostly good but I left comments and I have a few overall remarks:

• outside of one or two shared/helper modules, code in to_cmm should never directly build cmm expressions using the datatype constructors; everything should go through smart constructors (as if the cmm types were private)
• we need a better and clearer distinction for what belongs in the to_cmm_helper module. what I propose is the following:
  • have all smart constructors for cmm be in either cmm_helpers or to_cmm_helpers (and thus if you prefer to not have a to_cmm_helper file, move all constructors to the cmm_helpers file) (and again ideally all of these should in my opinion be in the cmm module directly, but it's probably too late for that)
  • have simple/arg_list functions that consume flambda2 things and produce cmm be shared be in a to_cmm_shared file

[Gbury]

I left a lot of small comments, which have to be addressed.

Some notes:

• the formatting of cmm_helpers destroyed a lot of formatting in comments which have to be restored, and some parts of the code building some large cmm expressions became (even more) unreadable
• we will really need a second PR to harmonize cmm_helpers.ml: put the basic smart constructors for cmm expression at the top, and re-use them throughout the file, and try and decide on a common pattern for function signatures (i.e. labelled arguments or not, debuginfo as first or last argument, etc...).

[mshinwell]

Regarding a second PR to deal with moving functions around in Cmm_helpers, I think we should wait. The desired situation is that Flambda 2 becomes the only middle-end, at which point a lot of the original Cmm generation code can be deleted. That seems like the right time to make further improvements to the file.

[Gbury]

That looks good, waiting for CI to be green.