it seems that, to be truly useful, pure fns really need the
ability to modify their parameters.  alternatively, we could
rewrite the functions that modify their arguments to take/return.

This reverts commit ec5cbb4f.

This reverts commit aabf84cd.

This reverts commit adb717b5, reversing
changes made to aabf84cd.

Incoming

Classes with no fields don't really make sense, so forbid them
(just as records with no fields aren't allowed). Closes #2509

There were bugs in the code path for initialization of non-spilled
locals, the code-path is hit approximately never, and mem2reg can do
it for us.
I think this will let us kill spill_map.

A class with a dtor has a different representation (a pair of a bit field
and a pointer to the class itself), and previously, the trans code
was selecting out the second field on any reference to a variable whose
type was a class with a dtor. This turned out to be wrong in the case where
a closure captures a variable with such a type. Changed the code to only
insert the field selection on a class field access.

get_item_val could generate an internal compiler error if a class with
a dtor was used before the class itself got translated, because it was
assuming that the dtor's node ID was always in the item_symbols table. Fixed
it to create a new symbol if necessary, and use it subsequently.