Rewrite processing of lines with "label":"contains" and "label":"item"
for appending for arrays instead override then when lines repeat few times
with equal outV.

RAM usage of parsing LSIF references are now back to normalcy and
feature works.

This removes the code in workhorse for determining whether parsing
references is enabled or not. It'll now always be enabled.

This is to improve RAM usage when parsing references utilizing the
existing `cache` struct.

Reduces memory usage when parsing LSIF file of gitlab-workhorse
from 7MB to 4MB.

Extract references parsing logic to a separate file and struct.
Also modified `Item` struct to have `Line` as `int32` so it'll
be fixed size.

This is in preparation for another MR wherein references will be
stored in file instead of memory.