fix: move analyzer regression with variants

[zerbina]

Summary

Separate the switch lowering and destructor injection into two proper
passes and apply them concurrently, fixing a move analyzer regression.

Details

• move the switch lowering out of injectDestructorCalls
• pass a Changeset instead of a mutable MirBody to the two passes
  in injectdestructors
• destructor injection and switch lowering are applied in a single
  batch by backends.process
• the switch lowering is still disabled for the VM backend
• update the doc comment of injectDestructorCalls to reflect reality

Move Analyzer

Applying both passes concurrently - instead of the destructor-
injection/move-analysis after switch lowering - fixes a regression
from ae54312

The move analyzer now sees a discriminator assignment as:

x.discr = y

instead of:

if x.discr != y:
  =destroy(x) # variant destructor
x.discr = y

allowing it to perform automatic moves from non-variant fields again
(refer to tmove_from_non_variant_field.nim).

Misc

• fix MIR pretty-printing for object/ref constructors -- the colon was
  missing between the field and value

[saem]

Separate the switch lowering and destructor injection into two proper passes and apply them concurrently, fixing a move analyzer regression.

Aren't they applied more sequentially, and not concurrently?

[saem]

One minor question about the PR message body, but otherwise it's good to go.

[zerbina]

Separate the switch lowering and destructor injection into two proper passes and apply them concurrently, fixing a move analyzer regression.

Aren't they applied more sequentially, and not concurrently?

Previously, both passes were implemented in the injectDestructoCalls procedure, with application working as follows:

input ----(`switch` lowering)----> X ----(destructor injection)----> output

now, application change to:

input ----(`switch` lowering)----> output
    \                              /
      ---(destructor injection)---

that is, both passes are applied to the same tree and the changes are then (effectively) merged. While it would have been possibly to keep injectDestructorCalls applying both passes, I separated them so that the decision-making of whether to actually apply the passes moves can move to the callsite.

[saem]

/merge

Thanks for the detailed explanation, the diagrams are super handy.

[github-actions]

Merge requested by: @saem

Contents after the first section break of the PR description has been removed and preserved below:

---

Notes for Reviewers

• follows up to js: support variant destructors #1267
• also a small preparation for mir: implement the new control-flow primitives #1281