cuda prefill failed on NVIDIA A100-SXM4-80GB during both inference and bench

**Description**
Hi, I am encountering a cuda prefill failed error when running both ds4 and ds4-bench. The CUDA backend initializes properly, and the model maps memory successfully, but it consistently fails immediately during the prefill stage.

**Environment**
GPU: NVIDIA A100-SXM4-80GB (sm_80)
Model: ds4-flash-imatrix.gguf (153.33 GiB)
Build Command: make cuda sm_80

**Steps to Reproduce**
1. Running ds4-bench:
```shell
(ds4) root@product-1-23:/workspace/Projects/codes/ds4# ./ds4-bench -m ./gguf/ds4-flash-imatrix.gguf --prompt-file speed-bench/promessi_sposi.txt --ctx-start 2048 --ctx-max 65536 --step-incr 2048 --gen-tokens 128
ds4: CUDA backend initialized on NVIDIA A100-SXM4-80GB (sm_80)
ds4: CUDA registered 153.33 GiB model mapping for device access

ds4: CUDA startup model cache prepared 153.32 GiB of tensor spans in 0.000s
ds4: cuda backend initialized for graph diagnostics
ds4-bench: context buffers 1742.43 MiB (ctx=65665, backend=cuda, prefill_chunk=4096, raw_kv_rows=4352, compressed_kv_rows=16418)
ctx_tokens,prefill_tokens,prefill_tps,gen_tokens,gen_tps,kvcache_bytes
ds4-bench: prefill to 2048 failed: cuda prefill failed
```

2. Running standard inference:
```shell
(ds4) root@product-1-23:/workspace/Projects/codes/ds4# ./ds4 -m ./gguf/ds4-flash-imatrix.gguf -p "Explain Redis streams in one paragraph."
ds4: CUDA backend initialized on NVIDIA A100-SXM4-80GB (sm_80)
ds4: CUDA registered 153.33 GiB model mapping for device access

ds4: CUDA startup model cache prepared 153.32 GiB of tensor spans in 0.000s
ds4: cuda backend initialized for graph diagnostics
ds4: context buffers 1053.75 MiB (ctx=32768, backend=cuda, prefill_chunk=4096, raw_kv_rows=4352, compressed_kv_rows=8194)
processing 16 input tokens: 16/16 (100.0%)
ds4: prompt processing failed: cuda prefill failed
```