This is a serious bug. But extremely hard to reproduce, so it went
undetected for 8 years. One needs two resurrections with different
main nodes, which are both in a hash chain which gets relinked on
key insertion where the colliding node is in a non-main position. Phew.
Thanks to lbeiming.
Important: this changes the semantics of the write barrier!
Carefully read the big comment block in lj_obj.h
This helps HREFK key slot specialization and allows safely hoisting
HREF/HREFK across GC steps, too (fix for a barely reproducible bug).
Dead keys are only removed during a table resize (as before).
Compile math.sinh(), math.cosh(), math.tanh() and math.random().
Compile various io.*() functions.
Drive the GC forward on string allocations in the parser.
Improve KNUM fuse vs. load heuristics.
Add abstract C call handling to IR.
