test(fuzz): assert the round-trip oracle can re-decode what it encodes

[Greg Lamberson (glamberson)]

Summary

The pdu_round_trip oracle discarded the result of its own re-decode, so an encoder
that emitted bytes it could not read back passed silently. This asserts it, fixes the
one pre-existing failure that turning it on exposes, and adds the auto-detect PDUs that
were missing from the type list.

Stacked on #1491. That PR fixes the BandwidthMeasureStop asymmetry which this oracle
would otherwise fail on immediately, so it needs to land first.

Why the assertion was missing, and why that did not work

The macro ended in let _ = decode(&encoded);. The doc comment was explicit that this
was deliberate:

Re-decode returning Err. Surfaces an asymmetry but not a memory-safety bug; tracked
via filed follow-up issues, not this oracle.

In practice they were not tracked. The BandwidthMeasureStop asymmetry in #1491 went
unnoticed and was found by hand, while writing an unrelated test. Emitting bytes we
cannot read back is a real defect on the wire, so it is asserted now.

What turning it on found

LogonInfoVersion1::encode clamps both names to their fixed-width fields with resize,
then wrote the untruncated length into domainNameSize and userNameSize. For a name
at the field boundary that advertises a size larger than the field, and this crate's own
decoder rejects domainNameSize > DOMAIN_NAME_SIZE_V1, so the encoder could produce a
PDU it could not read back. Fixed in the first commit here: the sizes now describe what
is actually emitted.

Found in under a minute of fuzzing.

What is deliberately not asserted

An earlier revision also asserted byte stability, that decode-then-encode reproduces the
input bytes. That is not a valid property here and the fuzzer said so within a minute.

Several decoders normalise. LogonInfoVersion1 reads domainNameSize, range-checks it,
and then keeps only the trimmed string; the size field is not represented in
LogonInfo. A PDU whose size field disagrees with its own zero padding therefore cannot
re-encode to identical bytes however correct both halves are. Asserting it reported
design as breakage, so it was removed. Only re-decodability is asserted.

Auto-detect coverage

AutoDetectReqPdu and AutoDetectRspPdu were absent from the type list. Their absence
is the other half of why the #1491 asymmetry survived: even with the assertion, nothing
was feeding those types through the oracle.

Verification

• 3.5M runs clean with the stack applied
• reverting either fix reproduces a failure within 45 seconds, so the assertion bites
• cargo xtask check fmt -v green
• cargo xtask check lints -v green
• cargo xtask check tests -v green
• cargo xtask check typos -v green
• cargo xtask check locks -v green
• cargo +nightly check --locked in fuzz/ green (that workspace's lock is not covered
  by xtask check locks)