mirflow: make optimization algorithm non-recursive

[zerbina]

Summary

Change the space optimization applied to the internal, structured
control-flow representation used in the code generator to use a non-
recursive approach.

Details

• reorder the fields in mirflow.Stmt in order to reduce the type's
  required space (24 byte vs. 32 byte, on a 64-bit architecture)
• apply the block folding right when emitting a block. This gets rid of
  the recursive post-processing
• implement a simple optimization pass for inlining some CGIR block
  continuations into dispatcher targets
• add a dedicated statement kind corresponding to the MIR's continue,
  so that the latter is visible as a terminator to statement analysis

The optimization pass eliminates deep nesting when there are case
statements with many branches, though it's unable to fully optimize the
hash-table lookup generated for case statements over string values.

Processing of the structure control-flow representation in mir2cg
still uses recursion, meaning that its still susceptible to stack
overflow when the structures are very nested.

---

Notes for Reviewers

• the first half of fixing thtmlparser.nim not compiling with a debug compiler

[zerbina]

Changing mir2cg to not use recursion for traversing the control-flow representation isn't as easy, given that there are multiple, mutually recursive routines. It's likely that the test (or other code with large string case statements/expression) won't compile for a long while (until mir2cg is replaced with something better).

As for the CI failure, the C code generator currently emits code that relies on C compiler extensions or C23 (i.e., declarations immediately following a label), and the clang used by the macOS CI seems to be rather strict. I'll make a separate PR for fixing this.