1. Fix hprof heap dump to include constant pool connections to classes
2. Find better way to provide trace information on "before" VM_INIT"
objects. Idea: Add jvmti phase to TraceKey, turn tracking on at VM_START,
do HeapIterates at VM_INIT to mark all objects with no tag.
3. Consider adding all extra information to a new xml output format? format=x?
(Compatibility issues changing the format=b or format=a output?)
Or put all new output on flags? Default on, compatibility concerns can
be addressed by turning the new output off?

Idea: Like demo/jvmti/mtrace... Use BCI on all new/newarray bytecodes to
just record the cnum/mnum before the new/newarray bytecode, that way we have
at least the location of the allocation? Could then consider cleaning up
the stack traces gotten at Object.<init> too?
If before Vm-INIT (or thred==NULL)

TLS table should allow for (env,thread==NULL) as a tls entry.
If thread==NULL (before VM-INIT), then you try and sample the current
cnum/mnum saved by the BCI at the new/newarray opcode?

Also need to capture JVMTI_EVENT_VM_ALLOC_OBJECT.