Very light speed improvements
    
    When comparing of two kernel trees showed that for huge changesets
    involving highly recursive types, leaf functions used by the
    categorization code dominate the performance profile.
    
    This patch introduces some changes that help gain around 30 seconds
    (out of 14 minutes) on a non-optimized build, when comparing 4.7 nd
    4.8 fedora kernels (between fedora 24 and 26).
    
    	* include/abg-comp-filter.h (has_harmless_name_change): Pass smart
    	pointers by reference.
    	* src/abg-comp-filter.cc (access_changed)
    	(function_name_changed_but_not_symbol)
    	(non_static_data_member_type_size_changed)
    	(static_data_member_type_size_changed, is_compatible_change)
    	(decl_name_changed, has_harmless_name_change):  Pass smart
    	pointers by reference.
    	* include/abg-ir.h (decl_base::set_context_rel): Take a bare
    	pointer, not a smart pointer.
    	* src/abg-ir.cc (decl_base::priv::context_): Make this data member
    	be a naked pointer, not a smart pointer.
    	(decl_base::priv::priv): Initialize it.
    	(decl_base::priv::~priv): New constructor.
    	(decl_base::{get_context_rel, set_scope}): Adjust.
    	(class_decl::method_decl::{method_decl, set_scope}): Likewise.
    	(equals): In the overload for var_decl, compare the type of the
    	var first as that might be faster (to detect var_decls with
    	different types) in the general case where types are
    	canonicalized.
    
    Signed-off-by: Dodji Seketeli <dodji@redhat.com>