cuda: fix uninitialized IQ2 dequant LUTs in batch MoE expert-tile kernels (n_embd > 4096)

[giannisanni]

Summary

The batch MoE expert-tile kernels (moe_gate_up_mid_expert_tile8_{row32,row2048,rowspan}) load the IQ2_XXS dequant tables (s_iq2_grid, s_iq2_signs) into shared memory inside the xq_blocks <= 16 activation-staging branch, but the dev_dot_iq2_xxs_q8_K_block8_deq_lut calls consume them unconditionally.

Models with n_embd <= 4096 (16 q8_K blocks, e.g. V4 Flash) always take the branch, so the bug is invisible there. Any model with n_embd > 4096 skips it and the entire batch gate/up dequant runs against uninitialized shared memory: multi-token batch encodes (prefill, MTP verify) produce fluent garbage at full speed, while single-token decode stays correct because the n_tokens == 1 path uses different kernels. That combination (fast, plausible-looking token stream, wrong content, decode fine) makes it easy to misattribute to model quality.

Fix

Hoist the two LUT loads out of the staging conditional in the three kernels and make the __syncthreads() barrier unconditional. The moe_gate_up_mid_decode_lut_qwarp32_kernel has the same textual pattern but its launcher already guards use_decode_lut_gate on xq_blocks <= 16, so it is left unchanged.

How it was found and tested

Found on a GLM-5.2 CUDA port (n_embd 7168 = 28 blocks) where long-prompt batch prefill output was corrupt while short prompts (token-major path) were fine. Bisected with per-stage kill switches down to the expert-tile gate/up kernels, then to the conditional LUT load.

Before/after on the same hardware (RTX 4060 Ti 16GB, SSD streaming, 600-token prompt, greedy):

	prefill	output
before	6.5 t/s	fluent garbage
after	6.5 t/s	correct (retrieval question about the prompt answered right)

Flash-class models are unaffected by the change (they take the same staging branch as before; the LUT loads just move a few lines up).

Context: this came out of the streaming experiments in #495; the same uninitialized LUTs also corrupted 2-token MTP verify logits there.