xshared: Fix for memleak in option merging with ebtables

xshared: Fix for memleak in option merging with ebtables

The crucial difference in ebtables is that all extensions are loaded up
front instead of while parsing -m/-j flags. Since this loading of all
extensions before every call to do_parse() is pointless overhead (cf.
ebtables-restore), other tools' mechanism of freeing all merged options
in xtables_free_opts() after handling each command and resetting
xt_params->opts at the start of the parser loop is problematic.

Fixed commit entailed a hack to defeat the xt_params->opts happening at
start of do_parse() by assigning to xt_params->orig_opts after loading
all extensions. This approach caused a memleak though since
xtables_free_opts() called from xtables_merge_options() will free the
opts pointer only if it differs from orig_opts.

Resolve this via a different approach which eliminates the
xt_params->opts reset at the start of do_parse():

Make xt_params->opts be NULL until the first extension is loaded. Option
merging in command_match() and command_jump() tolerates a NULL pointer
there after minimal adjustment. Deinit in xtables_free_opts() is already
fine as it (re)turns xt_params->opts to a NULL pointer. With do_parse()
expecting that and falling back to xt_params->orig_opts, no explicit
initialization is required anymore and thus ebtables' init is not
mangled by accident.

A critical part is that do_parse() checks xt_params->opts pointer upon
each call to getopt_long() as it may get assigned while parsing.

Fixes: 58d364c7120b5 ("ebtables: Use do_parse() from xshared")
Signed-off-by: Phil Sutter <phil@nwl.cc>