Optimize ZEND_CAST to avoid zval copies

Optimize ZEND_CAST to avoid zval copies

The scalar type casts IS_NULL, IS_BOOL, IS_LONG, IS_DOUBLE and
IS_STRING will no longer require a copy when casting.

A copy is now only made when casting to IS_ARRAY and IS_OBJECT, if
the type doesn't already match.

I tweaked the reference handling for the type-already-correct case
to DEREF the zval after that check. References require a copy anyway,
so they can go through the slow codepath.