Simplify AEAD ProcessBytes dispatch and tidy WipeKeyMaterial

[Xor-el]

Refactor the large, duplicative ProcessBytes methods of the AEAD modes, keeping every hot kernel and the byte-for-byte crypto output unchanged -- only the dispatch scaffolding around them is restructured.

GCM

• Collapse the near-mirror encrypt/decrypt whole-block dispatch (the 8-4-2-1 tiered staircase plus the repeated two-block advance) into a single RunTieredWholeBlocks(AForEncrypt, ..., AHoldBack) helper. The hold-back (0 on encrypt, the tag size on decrypt) unifies the two threshold families; the per-tier methods are unchanged. 243 -> 110 lines.

EAX

• Extract the buffer drain / residue-stash / decrypt-stitch boilerplate repeated across the four dispatch paths into DrainInto / StashResidue / StitchDecryptBlock0. The OMAC lookahead sequencing and fused-kernel logic are untouched. 367 -> 314 lines.

ChaCha20-Poly1305

• Stop rejecting a nil output buffer in DoFinal; rely on the output-length check like the other AEAD modes. In Pascal an empty result array is nil, so the old check wrongly failed a legitimate empty-plaintext decrypt.

All modes

• Drop the redundant "if F <> nil then" guards around the WipeKeyMaterial fills; the fill helper already no-ops on a nil (or empty) array.

Tests

• Add a chunked-streaming equivalence harness (chunk sizes across every buffer/tag boundary, both directions, out-of-place and in-place, with per-call Result accounting) that runs each mode under the kernel on/off toggle and the software engines, so every dispatch tier is exercised.