In order to desugar `dbg` in a pipeline we need to allow a bare `dbg`
node in desugaring and only report it as an error if the bare node
survives to the next step of canonicalization. This means we move the
error code out of `desugar_expr` and into `canonicalize_expr`. This is
much simpler to do now that these functions use the same `env` struct,
since previously we would have had to pass down extra args to
`canonicalize_expr`. Sharing the `env` struct means that we also don't
have to worry about calculating `line_info` more than once.
e.g. in the scenario of `[ A Str, B, C, D ]`, rather than having one tag be nullable, we want
to store the tag id in the pointer. This generally saves 8 bytes for
every allocation, and prevents an allocations for all tags except A.
This new flag determines whether we should introduce a new kind to
represent lambda sets, or whether lambdas should be erased. The latter
is not yet implemented.
Previously, we would drop uninhabited captures from lambda sets' runtime
representations, which meant sets like
```
[L1 {a: Str}, L2 {a: []}]
```
had runtime representation
```
{Str}
```
rather than
```
Union({Str}, {[]})
```
if we drop unreachable lambdas from the representation, then the
reachable lambdas are somewhat more efficient to compile (as there are
less material tag options), but the compiler complexity increases
because we must represent voided capture sets in the lambda set.
Even if a lambda has voided captures, we must specialize it, because
failing to do so opens us up to losing relevant specializations needed
later on. See 2f7020aa31 for a
previous occurence of that.
As such, simply keep voided layouts in place during lambda set
compilation. The optimizer should elide them anyway.
