feat: add IpNetwork::to_ipv6_mapped and to_canonical for dual-stack conversion

[3Hren]

Closes an API gap on the IpNetwork enum: Ipv4Network has to_ipv6_mapped() and Ipv6Network has to_ipv4_mapped(), but the enum itself had no way to move between address families.

The motivating use case is dual-stack comparison. On the enum, contains/intersects return false/None across families by design, so lifting both operands into IPv6 space is the way to compare them meaningfully.

API

• IpNetwork::to_ipv6_mapped(&self) -> Ipv6Network — embeds the network into IPv6 space. V4 delegates to Ipv4Network::to_ipv6_mapped(), V6 is the identity (the result is therefore not necessarily ::ffff:-prefixed).
• IpNetwork::to_canonical(&self) -> IpNetwork — mirrors core::net::IpAddr::to_canonical. V4 passes through, V6 collapses to V4 when it is IPv4-mapped.
• impl From<IpNetwork> for Ipv6Network, next to the existing From<Ipv4Network>.

Both methods are const fn, #[inline], #[must_use], and delegate to the existing conversion methods — no bit arithmetic is re-derived.

Ipv6Network::is_ipv4_mapped_ipv6 requires both the address to be ::ffff:-prefixed and the mask's top 96 bits to be all ones, so to_ipv4_mapped is the exact inverse of to_ipv6_mapped on the image of the embedding, and to_canonical never loses addresses.

Tests

12 unit tests and 9 property tests. Coverage includes non-contiguous masks in both families, the IPv4-compatible-but-not-mapped case, and the case where the address is ::ffff:-prefixed but the mask does not pin the top 96 bits (the network must stay V6 — collapsing it would drop addresses).

Two property tests pin the motivating use case: a.contains(&b) == a.to_ipv6_mapped().contains(&b.to_ipv6_mapped()), and likewise for intersects.

No benchmarks: these are two-arm delegating wrappers over already-benchmarked methods, so no new hot path is introduced.