drsuapi: fix DsGetNCChanges V6 parser to actually extract NTDS hashes

[psycep]

Summary

• Rewrites the hand-rolled DRS_MSG_GETCHGREPLY_V6 parser as a declarative NDR struct walker. The old parser drifted out of alignment within the first few REPLENTINFLIST entries, so secretsdump emitted zero NTDS hashes against any modern DC.
• Adds pkg/dcerpc/drsuapi/ndr.go with NDR Decoder primitives (alignment, conformant arrays, pointer referents, UTF-16LE strings, GUIDs, sticky error model).
• Adds pkg/dcerpc/drsuapi/getncchanges_v6.go with a parser built from MS-DRSR struct definitions.
• Removes ~950 lines of now-unreachable legacy V6 helpers from getncchanges.go.

Net: +603 / -956 lines.

Closes #9.

Correctness points the old parser got wrong

1. V6 fixed header is 148 bytes, not 136. The old code omitted cNumValues + rgValues + dwDRSError and treated those 12 bytes as the start of pNC's DSNAME.
2. DSNAME.StringName is a pure conformant array (NDRUniConformantArray), not conformant-varying. No Offset/ActualCount header between NameLen and the WCHAR elements.
3. UPTODATE_VECTOR is V2_EXT on current DCs: 32-byte cursors (UUID + USN + DSTIME), not the V1 24-byte cursors the old parser hard-coded.
4. REPLENTINFLIST serialization order: it's a linked list via pNextEntInf. NDR uses strict DFS through the first-encountered pointer in each struct, which lays out all N fixed parts consecutively, then the non-pNext deferreds (pName, pAttr, pParent, pMeta) unwind bottom-up (the deepest node's deferreds appear first, the head's last). The parser iterates captured headers in reverse to match.
5. PROPERTY_META_DATA_EXT_VECTOR alignment: the array's MaxCount is hoisted to the front by NDR early conformance, but hoisted MaxCount uses only its primitive (4-byte) alignment. The struct's 8-byte alignment applies after MaxCount, before cNumProps and the elements. Getting this wrong drifted every entry whose prior pParent UUID landed on a 4-aligned (non-8-aligned) position. This was the final bug blocking full-response parsing.

Verification

Live DC, compared to impacket-secretsdump:

$ diff impacket.txt gopacket.txt
< Guest:501:aad3b435b51404eeaad3b435b51404ee:31d6cfe0d16ae931b73c59d7e0c089c0:::

17/17 real password hashes match byte-for-byte (Administrator, krbtgt, vagrant, KINGSLANDING$, ESSOS$, NORTH$, and all 11 domain users). The single line Impacket emits that we don't is Guest:501 with the empty-password placeholder hash. Guest exists in the replication stream but has no unicodePwd attribute; Impacket emits a placeholder, gopacket filters to accounts where a real hash was extracted. Cosmetic output difference, no credential recovery gap.

Independent spec-driven Python reference walker (not Impacket; Impacket's own fromString on DRS_MSG_GETCHGREPLY_V6 doesn't deeply traverse the linked list): agrees with the Go parser on exactly the 17 unicodePwd-bearing DNs.

Offline parse of a 1MB captured response: 254 objects extracted from 307 replicated entries, end-to-end without drift.

Test plan

• go build ./...
• go vet ./...
• go test ./... (all existing tests pass)
• Live gopacket secretsdump -just-dc-ntlm against GOAD sevenkingdoms.local DC, all 17 hashes recovered
• Byte-for-byte diff vs impacket-secretsdump, identical on every real hash
• Regression for large multi-page dumps (fMoreData=1). Our test blob has fMoreData=0; this code path is unchanged from before but worth exercising against a bigger domain before closing any follow-on issues.