Stop re-validating addresses the decoder just generated

[EONRaider]

Target: 1.4.0 (Tier 1 — earn the benchmark)

Every header's __post_init__ runs the address validators, including
when the instance was built by decode(). So on the decode path:

1. bytes_to_mac() turns six raw bytes into "00:07:0d:af:f4:54".
2. __post_init__ immediately runs validate_mac_addr() on it, which
   matches that string against mac_regex.

The regex cannot fail — the string was mechanically generated from six
bytes moments earlier. The same applies to validate_ipv4_addr()
against socket.inet_ntop() output.

After #82 and #83 land, re.Pattern.match becomes the single largest
remaining cost in a corpus profile (27,680 calls in an 80-repetition
run).

Measured

300,000 calls: validate_mac_addr 0.229 s, validate_ipv4_addr
0.120 s → 2.3 µs per Ethernet+IPv4 frame spent proving something
already true.

What to do

Let decode() construct through a path that skips validation, while
keeping strict validation for user-supplied values. Strictness on
construction is a deliberate differentiator and must not be weakened —
the only claim here is that the decoder needn't re-check its own
output.

Options worth weighing: a private _unvalidated() classmethod using
object.__new__ plus object.__setattr__; or a module-private
sentinel that __post_init__ checks. The first avoids a per-instance
branch; the second is less invasive. Whichever is chosen, it must not
appear in the public API.

Acceptance criteria

• Ethernet(dst="nonsense", ...) still raises
  InvalidMACAddressError.
• Ethernet.decode(...) performs no regex matching.
• Round-trip and corpus tests unchanged and green.