GLM-DSA: fix -fa 0 garbage perplexity on the batch>8 indexer path (private seed, do not corrupt KQ_mask)

[mb8565]

Follow-up to the GLM-DSA work (#2045, #2066, #2067). With -fa 0 (no flash attention) and --dsa, the DSA path still produces garbage perplexity on the batch >8 (prefill) indexer path. -fa 1 --dsa is fine. #2067 already fixed the small-batch (<=8, TG/decode) path as a side effect of switching it to a single non-inplace ggml_add. This PR handles the remaining batch >8 path.

Cause: build_deepseek2_dsa_indexer seeds the indexer score accumulator with a view of KQ_mask. On -fa 0, KQ_mask is the raw F32 input tensor. The small-batch path (#2067) accumulates with a non-inplace ggml_add, so it does not touch KQ_mask. The batch >8 path still accumulates the heads with ggml_add_inplace, which writes into the shared KQ_mask buffer and corrupts the causal mask that build_deepseek2_dsa_sparse_mask and the later softmax layers read back. On -fa 1 the mask is F16, so the indexer's cast to F32 already gives a private buffer.

Fix: in the batch >8 branch, take a private copy of the F32 seed before the in-place accumulation, matching what the small-batch path and -fa 1 already do:

} else {
    // -fa 0: the seed still aliases KQ_mask; copy it so the in-place accumulation below does not
    // corrupt the shared causal mask. -fa 1's F16 mask was already cast to a private F32 buffer above.
    if (KQ_mask->type == GGML_TYPE_F32) {
        indexer_score = ggml_cont(ctx0, indexer_score);
    }
    for (int head = 0; head < n_ihead; ++head) {
        ...
    }
}

The small-batch / TG path is left untouched, so it keeps the #2067 performance work, and the copy is gated to -fa 0 (raw F32 mask) so -fa 1 prefill pays nothing. Cost is one n_kv x n_tokens F32 copy per full layer on the batch >8 -fa 0 path only.

Results (unsloth IQ2_M GGUF, top_k 2048, CPU unless noted; on current main dbe2ecb / #2067):

• -fa 0 --dsa 4K PPL (batch 512, the batch >8 path): thousands (broken) -> 2.7134 with the fix (dense 2.6972, -fa 1 --dsa 2.7111)
• -fa 0 --dsa 4K PPL on 3x P100 (CUDA, --cpu-moe): 2.6980, sane and in the dense / -fa 1 band (not garbage), so the fix works on CUDA (pre-GLM-DSA: minor optimization #2066 tree, same copy fix)
• Dense-equivalence oracle (DSA top_k >= n_kv must == dense, chunk-for-chunk): bit-exact on both -fa 1 and -fa 0-with-fix -> the attention machinery is correct, the DSA delta is pure key selection (pre-GLM-DSA: minor optimization #2066 tree, same copy fix)
• Coherence: -fa 0 clean at 9K (CPU), coherent at 24K (GPU, 21272-token prompt); -fa 1 unaffected throughout (pre-GLM-DSA: minor optimization #2066 tree, same copy fix)
• Note: -fa 0 --dsa at very long context is memory-heavy (per-layer sparse mask); 24K is the max on 48GB VRAM

Performance: decode is unaffected. The #2067 small-batch fast path is batch-gated, not FA-gated, so -fa 0 decode also takes it and never reaches the patched batch >8 loop. The fix adds one n_kv x n_tokens F32 copy of the causal-mask seed per full layer on the -fa 0 prefill path, small next to the per-head matmul loop it guards, and it leaves the TG compute-buffer path that #2067 optimized untouched. If you would rather fold the copy into the accumulation, making head 0's add non-inplace and keeping the rest in-place is an equivalent, slightly cheaper alternative to the ggml_cont.

[mb8565]

Noticed #2068 is in flight on the same file. This one only touches the batch >8 indexer accumulation, which #2068 leaves alone, so they are independent and it should rebase cleanly on top. Happy to rebase once #2068 lands.