PowerPC TLS miscounting PLT for __tls_get_addr

ppc*_elf_tls_optimize decrements the PLT refcount for __tls_get_addr
when a GD or LD sequence can be optimized.  Without tls marker relocs
this must be done when processing the argument setup relocations.
With marker relocs it's better done when processing the marker reloc.
But don't count them both ways.

Seen as "unresolvable R_PPC_REL24 relocation against symbol
`__tls_get_addr_opt'" (and other branch relocs).

	* elf32-ppc.c (ppc_elf_tls_optimize): Don't process R_PPC_TLSLD
	with non-local symbol.  Don't double count __tls_get_addr calls
	with marker relocs.
	* elf64-ppc.c (ppc64_elf_tls_optimize): Likewise.