feat(net): apply scheme-less --net-allow/--net-deny rules to both TCP and UDP

[congwang-mk]

Closes #132.

A --net-allow / --net-deny spec with no scheme now applies to both TCP and UDP: it expands to one rule per protocol at parse time. tcp://, udp://, and icmp:// keep pinning a single protocol, so nothing becomes inexpressible; ICMP is never implied and still requires an explicit icmp:// rule.

Motivation (from the #75 discussion):

• --net-allow '*' --net-allow 'udp://*' collapses to --net-allow '*', and --net-allow 1.1.1.1:53 now actually permits ordinary (UDP) DNS queries.
• The deny side is the security-relevant half: --net-deny 169.254.169.254 previously blocked only TCP to the metadata service, silently leaving the UDP path open when UDP was allowed. A scheme-less deny now closes both.

Implementation notes:

• Expansion happens in NetRule::parse_spec, which now returns Vec<NetRule> (one rule for scheme'd specs, a TCP + UDP pair otherwise). Every rule still carries exactly one concrete protocol, so resolution, seccomp routing, the Landlock TCP fast path, and port remap are unchanged.
• The serde default of protocol: tcp on NetRule is removed; the field is always written explicitly, and pre-1.0 policies are not kept compatible by policy.
• format_net_rule (profile round-trip) now always renders the scheme: a bare rendering would re-widen an explicit tcp:// rule on reparse.
• resolve_net_allow resolves each unique hostname once and shares the IP set across protocol passes; per-pass lookups could disagree under DNS round-robin and duplicated the synthetic /etc/hosts lines.
• README, sandbox-reference, and the Python SDK docstrings are updated to the new default.

Verified end to end with the built CLI: UDP sendto succeeds under --net-allow '*', stays blocked under --net-allow 'tcp://*', and a scheme-less --net-deny 127.0.0.0/8 now blocks UDP to the CIDR while leaving other destinations open. Full Rust workspace suite (18 suites) and the Python suite (386 tests) pass.