Fix/homa destroy ordering

[acbk2b]

What?

Shared Homa state destroyed before namespace cleanup

• On registration failure, homa_destroy() ran before unregister_pernet_subsys().
• Unregistering invokes homa_net_exit() / homa_net_destroy(), which still require the shared Homa state.
• Move homa_destroy() after the other unregister/cleanup operations.

Timetrace /proc state left behind on load failure

• The error path did not call tt_destroy(), leaving stale proc entries and causing a delayed crash when accessed.
• Call tt_destroy() in the error cleanup path, guarded by #ifndef __UPSTREAM__ because timetrace is excluded from the upstream kernel tree by util/strip.py.

tt_init() failure ignored

• homa_load() did not check the result of tt_init().
• Return immediately if tt_init() fails; no downstream Homa initialization has occurred, so the general cleanup path is unnecessary.

Also fixed a small bug in homa_rpc_ack__multiple_acks which attempted to destroy an uninitialized socket

Why?

The above bugs, combined with missing .netns_ok caused the Homa module to fail to load + crash my RHEL 8 test box.
.netns_ok is only applicable to RHEL8, however these changes are generally applicable to the module and are not exclusive to RHEL8

Each bug fix addressed a different failure mode the occurred during my testing. Stacking all of these fixes allowed me to successfully load the kernel module.

Testing?

I have added unit tests covering each bug to ensure a future regression would be detected

Anything else?

These changes will need to be back-merged into rhel8 as well