glm: keep rope models on the indexed causal attention kernel (fixes #932)

[garnetlyx]

Summary

b0c31af ("Improve GLM 5.3 attention memory and batching") added a dense
compact flash-attention prefill path
(ds4_gpu_glm_attention_dense_compact_lora_causal_tensor) that scores the
shared latent cache directly with no separate RoPE contribution. That is only
exact for models whose DSA cache row has no RoPE tail — GLM-5.3 Flash
(n_rot == 0, "Raw DSA cache row: 512 latent dimensions and no RoPE tail").

On GLM-5.2 and full GLM-5.3 (n_rot == 64), the rotated RoPE term from
k_rope_cache is silently dropped. Every prompt with at least
glm_graph_flash_attention_prefill_min_tokens() (= 24) tokens gets wrong
attention, and generation degenerates into fluent but unrelated text — this is
issue #932.

Evidence

• Two independent git bisect runs (mine and GLM 5.2 Q4 on Metal (M3 Ultra) produces fluent but unrelated output after commit b0c31af #932's) land on b0c31af with
  4771329 as the last good parent.

• Threshold matches the gate exactly: 21-token prompts behave, 25-token prompts
  produce garbage (glm_graph_flash_attention_prefill_min_tokens() == 24).

• Repro on M3 Ultra (Metal), GLM-5.2-UD-Q2_K_RoutedQ2K.gguf:

  ./ds4 -m gguf/GLM-5.2-UD-Q2_K_RoutedQ2K.gguf --ctx 4096 --tokens 240 \
        --temp 0 -p "Was ist 2+2? Antworte nur mit der Zahl."
  # -> unrelated fluent text (differs run to run at temp 0), never "4"

• The indexed kernel the old path used documents the invariant in
  ds4.c (glm_graph_forward_indexed_tokens): "The compact causal score
  includes both absorbed MLA and the separate RoPE query/key contribution" —
  it passes layer_k_rope_cache + DS4_N_ROT + rope freq params. The dense
  flash path passes only layer_kv_lora_cache.

Fix

Gate glm_graph_use_dense_compact_attention_prefill() on DS4_N_ROT == 0, so
rope models keep the indexed causal kernel. GLM-5.3 Flash (n_rot == 0)
behavior is unchanged; no kernel code is touched.

Verification

M3 Ultra (Metal), GLM-5.2-UD-Q2_K_RoutedQ2K:

• The repro above now answers 4, byte-identical across runs (determinism
  restored at temp 0).
• Prompt-length ladder (17/21/25/29/33/81/273 tokens, temp 0): every length
  returns exactly LENGTH_AB_OK.
• DS4_GLM_DISABLE_FLASH_PREFILL=1 was the interim workaround; this fix is
  verified equivalent without it.

Not verified locally: GLM-5.3 Flash regression (no Flash GGUF on this host);
the gate keeps its behavior identical by construction.

Fixes #932