Fix autostart-on-host-failure tag never matching, load tags eagerly

[Brittlejf]

Summary

PR #2 added opt-in autostart of instances tagged
autostart_instance_on_host_failure after a hypervisor reboot, but the
tag check never matched, so the feature did the opposite of its intent.
This fixes the check and avoids a per-instance DB query at startup.

Background

_init_instance decides whether to resume a guest on host boot:

tag_state = False
if "autostart_instance_on_host_failure" in instance.tags:
    tag_state = True
expect_running = (db_state == power_state.RUNNING and
                  drv_state != db_state and tag_state)

instance.tags is a TagList of Tag objects, not strings, so
"..." in instance.tags compares the string against Tag objects and
never matches. tag_state was therefore always False, making
expect_running always False — so no instance was ever resumed
after a host reboot, including the tagged ones the feature is meant to
autostart.

Changes

• Fix the tag check — compare against the tag names
  ([tag.tag for tag in instance.tags]) instead of the Tag objects.
• Load tags eagerly — init_host now passes 'tags' in
  expected_attrs to InstanceList.get_by_host, so tags are joined in
  up front rather than lazy-loaded once per instance during
  nova-compute startup.
• Tests — add unit tests for the tagged (resumes) and untagged
  (does not resume) cases, and update the existing failed-resume test to
  tag its instance so it still reaches the resume path now that resume is
  gated on the tag.

Testing

nova.tests.unit.compute.test_compute_mgr resume tests pass under
stestr; flake8 is clean on the changed files.

[Brittlejf]

How 'tags' in expected_attrs affects servers with no tags set

Untagged servers are unaffected — they get an empty TagList, not an error or an extra query. Walking the load path:

1. DB layer (nova/db/main/api.py): instance_get_all_by_host → _instance_get_all_query. tags is not a "manual join" column (only metadata/system_metadata/pci_devices/fault are), so it falls into the else branch and is loaded via orm.joinedload('tags').
2. joinedload on a one-to-many relationship is a LEFT OUTER JOIN, so an instance with no tags is still returned — just with zero joined tag rows. db_inst['tags'] comes back as an empty list.
3. Object layer (nova/objects/instance.py:429): because 'tags' is in expected_attrs, _from_db_object runs obj_make_list(...) over that empty list and sets instance['tags'] to an empty TagList.

So for a server with no tags:

	Without 'tags' in expected_attrs (before)	With 'tags' (this PR)
instance.tags after load	not set → first access lazy-loads from DB (_load_tags, 1 query)	already set to empty TagList
Extra query at startup	yes, one per instance	none
"..." in [t.tag for t in instance.tags]	False (after the lazy query)	False (no query)

Net effect: identical behavior for untagged servers — tag_state is False either way, so they are not autostarted. The only difference is where the empty tag list comes from: an up-front outer join instead of a per-instance lazy query.

One caveat worth noting: joinedload is a real JOIN, so for instances that do have many tags it multiplies rows in the result set. Tags are short and capped (MAX_TAG_LENGTH plus a per-instance tag-count limit), so this stays in line with how nova already uses joinedload for relationships like security_groups.