There were more problems in gij related to the lack of locking when
resolving constants.  This patch fixes the problems by using
fine-grained locking whenever accessing the constant pool.

This lack of synchronization did not affect precompiled code because
gcj resolves constants by emitting calls _Jv_ResolvePoolEntry, and
that function locks the class whenever updating the constant pool.
While coarse, this does guarantee that the data races in the
interpreter cannot occur, which is why we've never seen this problem
with precompiled code.

In theory the class lock in _Jv_ResolvePoolEntry could now be removed
because the new fine-grained locking makes it unnececessary.  However,
I'm not going to remove it yet.

Andrew.