pack: extend expert packing to MQ4G256V2 (qt=44) and MQ4CG256 (qt=45)

[ghazni101]

Problem

try_load_packed_mq4_experts gates on MQ4_G256_QUANT_TYPE = 13, so MQ4G256V2 (qt=44) and MQ4CG256 (qt=45) experts fall through to per-expert loading: 20,480 individual hipMalloc calls instead of 80 packed blobs. On gfx1100 (24 GiB VRAM), the amdgpu driver overhead for 20,400 extra buffer objects costs 4.36 GB of VRAM.

All three formats share the same 136 B/group stride — only the 8-byte group header interpretation differs, which is handled by kernel dispatch on gpu_dtype.

Fix

Replace the constant gate with packable_mq4_dtype() accepting qt=13/44/45. Track the uniform dtype across experts and propagate it to WeightTensor views so kernel dispatch selects the correct V2/C kernels.

qt=30 (MQ4G256Lloyd, 160 B/group) and qt=47 (MQ6G256V2, 200 B/group) remain excluded — different strides.

Measurements

RX 7900 XTX (gfx1100), Ornith-1.5-35B-A3B MQ4R (qt=44), max_seq=512, Q8 KV, MTP off:

Metric	Before	After
VRAM after load	23.81 GB	19.45 GB
hipMalloc calls	20,872	472
Load time	38.0 s	2.8 s
Decode speed	224 tok/s	224 tok/s

Decode correctness verified (math + factual queries). V1 (qt=13) no regression.

Closes #744

[Kaden-Schutt]

Landed on beta at cc93d01bc (merge of b1955f4e2), validated on RX 7900 XTX with ornith-1.5-35B-A3B .mq4r and .mq4 (all 20,480 expert tensors qt=44), Q8 KV, greedy:

• VRAM after load: 24.45 → 20.10 GB (mq4r), 24.68 → 20.31 GB (mq4) — your −4.36 GB reproduces exactly.
• routed MQ4 expert packing: 512 per-expert weight buffers -> 2 layer blobs on every candidate load; absent on baseline.
• All 10 greedy battery answers byte-identical packed vs unpacked on both files; serve battery 5/5 + chain 5/5 on mq4r.
• Decode unchanged: 240.9 → 241.1 tok/s (mq4r), 169.2 → 169.7 (mq4).
• Load time: the 38 s → 2.8 s figure did not reproduce here — warm-page-cache sweep 3.70 → 2.97 s; cold loads are disk-bound on this host. The VRAM/alloc-count win is the durable part.

Notes for the record: qt=45 is included in the packable set but expert packing is RDNA3-only and qt=45 GEMM is gfx12-only, so that arm is unexercised; qt=13 path is unchanged by inspection (no qt=13 expert artifacts exist here to run). Will close with attribution once #730 promotes beta.