…e output in a temporary directory in the store

Puts the temporary FOD output copies in a temporary directory inside the
store instead of the (for Linux sandboxed builds) chroot.
This prevents file overwrite due to symlink following that std::filesystem::copy_file does.

Also applies the same output copying approach for impure derivations that don't
have network sandboxing and thus are subject to FD smuggling.

Fixes GHSA-g3g9-5vj6-r3gj.

(cherry picked from commit a760af8)

…ew enough kernels

This partially fixes the issue with cooperating processes being able
to communicate via abstract sockets. The fix is partial, because processes
outside the landlock domain of the sandboxed process can still connect to
a socket created by the FOD. There's no equivalent way of restricting inbound
connections. This closes the gap when there's no cooperating process on the host
(i.e. 2 separate FODs).

>= 6.12 kernel is widespread enough (NixOS 25.11 ships it by
default) that we have no reason not to apply this hardening, even though
it's incomplete.

ca-fd-leak test exercises this exact code path and now the smuggling
process fails with (on new enough kernels that have landlock support enabled):

vm-test-run-ca-fd-leak> machine # sandbox setup: applied landlock sandboxing
vm-test-run-ca-fd-leak> machine # building '/nix/store/s7brgi6pdr5f3n8yqlgmdlz8blb89njc-smuggled.drv'...
vm-test-run-ca-fd-leak> machine # building derivation '/nix/store/s7brgi6pdr5f3n8yqlgmdlz8blb89njc-smuggled.drv': woken up
vm-test-run-ca-fd-leak> machine # connect: Operation not permitted
vm-test-run-ca-fd-leak> machine # sendmsg: Socket not connected

(cherry picked from commit 44017ca)