fix(proto): canonicalize FourTuple's remote/local_ip at construction (noq#738)

[cuzic]

Closes #738. Alternative to #783 — see that PR for the minimal one-line fix;
this one takes a broader approach to the same root cause.

Problem

4bae6edd (the hotfix posted in #738) and #783 both patch individual
==/!= comparison sites in Connection::early_discard_packet to
canonicalize IPv4-mapped-IPv6 addresses (::ffff:a.b.c.d) before
comparing, since dual-stack sockets can report the same peer in either
form depending on the code path (observed on Android when a
physical-interface-bound socket is handed to noq via an abstract socket).

Patching comparisons one at a time only closes the specific sites someone
has actually hit a bug in. FourTuple derives PartialEq/Eq/Hash,
and it's used directly (not just via the two early_discard_packet sites)
in:

• Endpoint's HashMap<FourTuple, ConnectionHandle> connection-routing
  table
• Connection::is_probably_same_path
• PATH_CHALLENGE on-path detection
• OBSERVED_ADDRESS path matching
• peer migration detection

Every one of these has the same latent bug: two FourTuples that are
"the same address" from the OS's perspective but differ in
mapped-vs-plain representation will compare unequal.

Fix

Canonicalize remote and local_ip once, in FourTuple::new() itself,
so every downstream comparison/hash sees one canonical representation
regardless of which form the OS happened to report a given datagram's
address in. This makes the two early_discard_packet-specific patches
(4bae6edd, #783) unnecessary — they're subsumed by this fix — though
either could still land independently since they're harmless on their
own.

Not claiming this is strictly the better fix — a single well-placed
comparison-time fix (#783) is simpler to reason about and reviews faster.
Filing both so the maintainers can pick.

Testing

Same regression test and full-suite run as #783 (this PR includes the
same test-adaptation commit, since both branches were cut from 4bae6edd
independently):

• noq-proto/src/tests/multipath.rs::open_path_with_explicit_local_ip
  (from the noq-738 branch) passes; fails on unpatched main.
• cargo test -p noq-proto: 388 passed, 0 failed.
• Verified on a real Android device, same scenario as fix(proto): canonicalize remote IP in early_discard_packet's peer check (noq#738) #783 — Secondary
  established, then PhysicalWifi/PhysicalCellular both validate on
  the first attempt:

  opening path "physical-wifi" -> <redacted>:45823 (local_ip=Some(192.168.10.80), attempt 1/3)
  path "physical-wifi" established: id=PathId(2)
  opening path "physical-cellular" -> <redacted>:45823 (local_ip=Some(10.209.235.90), attempt 1/3)
  path "physical-cellular" established: id=PathId(3)