Barrier the pointer stores of a cppia field set by name

[Tutez64]

With HXCPP_GC_GENERATIONAL, a value set by name on a cppia instance can be freed while the instance still points at it.

Reflect.setField / Reflect.setProperty on a cppia object end in CppiaClassInfo::setField, which stores a member variable through CppiaVar::setValue. That function writes object and string pointers with a plain assignment, without a write barrier. The JIT and interpreted field setters (MemReferenceSetter) and FieldMapSet already have one.

When the instance is old and the value young, the value is then reachable only from the old instance. A generational collection does not traverse it, so the next one frees the value, and a later read of the field reads reused memory. In practice, a closure in a loaded script that reassigned an array field (compiled to a set by name) crashed on the next read of that array after a minor collection, or during a later full collection that followed the dangling pointer.

A standalone reproduction is attached: a scriptable host with HXCPP_GC_GENERATIONAL loads a small ASCII cppia script, lets the script's objects grow old with full collections, then has a closure set an array field through Reflect.setField and another read it, with only generational collections in between. On master the first read crashes, every run; with this change all five rounds pass. hxcpp-cppia-set-field-barrier.zip

The fix adds HX_OBJ_WB_GET after the fsString store and after every fsObject store (each case stores one object pointer at the field's offset), under HXCPP_GC_GENERATIONAL.

[Tutez64]

Added a test to test/cppia in 68322a0. Its -jit run also needs #1411, which fixes an unrelated JIT nursery-alignment bug the test exposed.