xtensa: improve vmlinux.lds.S sed post-processing

Current sed script makes assumptions about the structure of rules that
group .text sections in the vmlinux linker script. These assumptions
get broken occasionally, e.g.: 779c88c9 "ARM: 8321/1: asm-generic:
introduce.text.fixup input section", or 9bebe9e5 "kbuild: Fix
.text.unlikely placement".

Rewrite sed rules so that they don't depend on number/arrangement of text
sections in *(...) blocks.

Signed-off-by: Max Filippov <jcmvbkbc@gmail.com>