metal: default-on nr1=2 multi-column Q2_0 matvec (spec-decode verify)

[bri-prism]

What

Make the nr1=2 multi-column Q2_0 matvec the default for ne11 >= 2 (the small-batch / spec-decode verify path), mirroring the Q1_0 path — instead of requiring the opt-in GGML_METAL_Q2_0_NR1 env var.

Why

The nr1=2 variant reads each streamed q2_0 weight group once per 2 src1 columns, versus the mul_mv_ext route that re-streams per column. In-code microbench (M5 Pro): nr1_2 = 93.2 µs vs 122 µs ext at ne11=2 (+31%); ne11=4 via 2 passes 171 vs 183. Q1_0 already defaults to this; this brings Q2_0 (ternary targets) to parity — it helps the DSpark verify path (which runs the target at ne11 = draft depth).

It was previously opt-in only because of the nr1_3 occupancy cliff at ne11==3 (195 vs 152 µs ext, tpb=16). This change carves out ne11==3 (stays on ext), so only the measured-win widths engage.

How

• ne11 >= 2 && ne11 != 3 → nr1=2; ne11==3 → unchanged (ext); ne11==1 (base decode) → unchanged.
• GGML_METAL_Q2_0_NR1=1 restores the old routing; =2/3/4 still force as before.
• Output-identical — pure matvec routing, no numerics change.

Testing

• Ternary base decode (llama-bench tg128, ne11=1): 27.06 → 27.13 tok/s = unchanged (change correctly scoped to multi-column).
• DSpark verify (tlg6l1 × ternary-cont6k, k=4), default settings now auto-engages nr1=2: runs clean, accept 0.757 / τ 4.03, output correct.