v1.0.8

Escalate to SIGKILL when a provider CLI ignores SIGTERM

The previous fix (SIGTERM handler so Python's finally block actually
runs) turned out to be necessary but not sufficient: a fully-authenticated
agy keeps its own background goroutine loop alive (quota/experiment/model
refresh against Google's backend), and observed instances kept retrying
that loop every ~1s for days after the wrapper sent SIGTERM — the process
just doesn't reliably die from it once live. Verified with a process that
explicitly ignores SIGTERM (signal.SIG_IGN): the old cleanup left it
running forever; escalating to SIGKILL after a 2s grace period kills it
immediately, since SIGKILL can't be caught, blocked, or ignored. Applied
to all four wrappers. Bump to 1.0.8.