Teardown must not decide a test's verdict

[xroche]

Under set -e a failing command in an EXIT trap becomes the script's exit status, so a hiccup while tearing down fixtures fails a test whose assertions all passed. That is what turned 57_local-proxy-connect.test red on the Windows x64 leg of #765: five OK lines, no FAIL, exit 1. Every EXIT trap in the suite now runs teardown with errexit off, and the signal traps keep their own trap line, since sharing set +e with HUP/INT/QUIT/PIPE/TERM would leave errexit off for the rest of a signalled run and let a torn-down test still report success.

A || true on the rm would have been smaller, but it throws away the only diagnostic, and the evidence does not say which teardown command failed: a blocked rm -rf on the Windows runner exits 1 and prints "Device or resource busy", while the log shows exit 1 and nothing at all. The sharing violation in the issue is the plausible mechanism rather than a confirmed one, so whatever it really is now prints its own error.

99_teardown-status.test pins the semantics both ways and scans the suite so a new test cannot reintroduce the shape, the leaky combined trap included. The return 0 that three cleanup() bodies ended with never protected anything, since errexit fires at the failing command before it is reached.

Closes #773