opencl: route large narrow-hidden q6_K lm_head to the flat GEMV

[wanghqc]

Overview

The problem:

• The opencl backend has two q6_K GEMV kernels: the fast flat one and the slower gemv_noshuffle fallback.
• The rule that picks between them requires the weight to be big in both dimensions (ne[1] >= 32768 && ne[0] >= 2048).
• gemma-4 / gemma3n E2B ties its token embedding as the lm_head: a Q6_K weight of [1536 × 262144] — 330 MiB. It is huge, but its first dimension is 1536 < 2048, so the rule rejects it and it runs on the slow kernel every generated token:

Solution:

• Keep the existing rule exactly as it is, and OR-in one extra way to qualify: total weight size ≥ 256 MiB.
• One exception: the escape is skipped on the A7X (740), whose compiler miscompiles the flat q6_K kernel (wrong answers).
• Nothing changes for any weight the old rule already accepted, on any device.

Additional information

Performance on the X2-90 for gemma-4 E2B:

• Per-kernel is now ~23 GB/s => ~56 GB/s
• End to end (e2e) TG128 is now 27.62 → 38.58 = +39.7%

Requirements

• I have read and agree with the contributing guidelines Yes
• AI usage disclosure: Yes, prototyping.