Delete unused local allocation regions in Flambda 2

[mshinwell]

Flambda 2 can't currently delete unused Begin_region and End_region constructs (unlike the equivalent in Flambda 1), which I suspect is leaving some amount of unnecessary code around, together with inhibiting code motion in To_cmm.

This patch aims to rectify this. It equips the primitives that allocate (and Set_of_closures) with region variables identifying which region is to be allocated in (this will always be the most recent by construction, and will remain that way throughout the pipeline). Functions now take a my_region parameter which is substituted for the current region upon inlining.

Begin_region is now marked as having only generative effects rather than arbitrary effects, which still inhibits code motion in To_cmm (until such time as we figure out how to do that safely for these constructs), but allows deletion when the region variable it returns is unused. End_region remains as having arbitrary effects but a special case is added (in Simplify_let_expr and Expr_builder) to do two things. Firstly for the purposes of Data_flow the region variable in End_region is not counted as a use. Secondly when rebuilding a Let binding an End_region we consult Data_flow to see if the region variable (the argument to End_region) was actually used. If it was unused, then the End_region is dropped, which will in turn cause the Begin_region to be cleaned up without further ado.

There is one outstanding problem relating to the region stack during CPS conversion (Lambda_to_flambda) not containing the try_region variable. Currently there is a hack in place so I can try some examples but this needs to be fixed properly. There are also a few things to fix in Fexpr.

[mshinwell]

As a note I'm not planning on implementing full support for these enhancements in Fexpr yet. We will need to decide what to do about the maintenance of that code and update it to match the main Flambda 2 code in due course (there are also some other areas of divergence).

[mshinwell]

This PR also now aims to ensure that we don't end up with Local alloc modes appearing in terms, likewise no Begin_region or End_region primitives, when stack allocation is disabled.

[lthls]

I've left a number of comments, some of them applying to code that wasn't modified by this PR.
The only one that I think is serious is the one in to_cmm_env.ml about the extra binding for my_region.