feat(GFX1100-TG200): T21 keep-quant for V-head row-permuted GDN projections

[ghazni101]

Row: BACKEND-ROCM

Reconstruct the useful T21 leaf from external pull request #2804 on current
main: keep only the row-permuted Qwen3.5 GDN attn_qkv.weight and
attn_gate.weight tensors in their supported GGUF block encoding, permute
whole encoded rows at load time, and feed the existing vt::MatmulBT
keep-quant production path.

The accepted scope is opt-in and exact: VT_GDN_ROWPERM_KEEP_QUANT=1 enables
the route; unset, 0, malformed values, unsupported encodings, and
column-permuted GDN weights retain the current expansion behavior. The load
must own mutable bytes before permutation and must not mutate an mmap-backed
source.

The change needs a committed spec before implementation, red-first loader tests
through the production Qwen3.5 GGUF entry point, byte-level row-order evidence
for every admitted encoding, unchanged expansion controls, a production
forward/reachability witness, and fresh mutation review. Historical benchmark
and token claims from the stacked branch are not current acceptance evidence.

Source attribution: Ghazni's T21 commit
aa6f4ae6037e150d3756442c280d06790bb8dcd8 and its default-off correction in
a0b7333f7b62e3897a134c932b07f83aba0ffe78.