The problem here was that when using ObjectSpace (or object_id,
or FFI) we need to insert some hidden state into objects. This
state goes into the variable table, but does not get marshaled.
However, if we had both reserved space for instance variables and
for one of these hidden variables, we always would proceed to
marshal the object as though it might have instance variables. If
none of the non-hidden variables had been set, we marshaled the
object as having instance variables, but of zero length. This
broke several marshal specs, since objects that would otherwise
have no instance variable were marshaled as such.

This fix uses a second check against the variable list to ensure
that there are real variables to marshal (or an encoding variable)
before proceeding to emit the instance variable sigil and list.

This is not directly used other than to provide a more accurate
guess as to whether there are any non-hidden variable in use by
a given type. 0b6e3c0 fixes the original issue (marshal dumping
zero-length instance var lists).

The accessors of the lazy fields are now deprecated. Use the read
or write calls to get an accessor for the hidden fields.

In MRI most of these types are derived from the internal DATA
class.

We would like to have the table consider encoding, but until we
only use ByteList for identifiers this is not possible. We still
use Java's String in many places, in which case we only have
access to the raw bytes and not the encoding they should be
associated with, so the String-based lookup would never be able
to add this additional mask.

Unfortunately removing this additional hash dimension breaks one
Symbol marshal spec, but it was broken before this work as well.

This was the cause of Kernel#eval regressions in jruby#4576.