[Attention] Abstract the MLA prefill backends and eliminate cuDNN

docs/design/attention_backends.md

@@ -189,10 +189,10 @@ configuration.
 
 | Backend | Description | Compute Cap. | Enable | Disable | Notes |
 |---------|-------------|--------------|--------|---------|-------|
-| TRT-LLM Ragged‡ | TensorRT-LLM ragged attention | 10.x | Default on SM100 | `-ac.use_trtllm_ragged_deepseek_prefill=0` | DeepSeek R1 dims only |
-| FlashInfer | FlashInfer CUTLASS backend | 10.x | `-ac.disable_flashinfer_prefill=0` | `-ac.disable_flashinfer_prefill=1` | DeepSeek R1 dims only |
-| cuDNN | cuDNN-based attention | 10.x | `-ac.use_cudnn_prefill=1` | `-ac.use_cudnn_prefill=0` |  |
-| FlashAttention | FlashAttention varlen (FA2/FA3) | Any | Default fallback | Use other backends | FA3 on SM90, FA2 otherwise |
+| `TRTLLM_RAGGED_PREFILL`‡ | TensorRT-LLM ragged attention | 10.x | `-ac.mla_prefill_backend=TRTLLM_RAGGED` | `-ac.mla_prefill_backend=FLASH_ATTN` | DeepSeek R1 dims only |
+| `FLASHINFER_PREFILL` | FlashInfer CUTLASS backend | 10.x | `-ac.mla_prefill_backend=FLASHINFER` | `-ac.mla_prefill_backend=FLASH_ATTN` | DeepSeek R1 dims only |
+| `CUDNN_PREFILL` | cuDNN-based attention | 10.x | `-ac.mla_prefill_backend=CUDNN` | `-ac.mla_prefill_backend=FLASH_ATTN` | DeepSeek R1 dims only |
+| `FLASH_ATTN_PREFILL` | FlashAttention varlen (FA2/FA3) | Any | Default fallback | Use other backends | FA3 on SM90, FA2 otherwise |
 
 > **‡** TRT-LLM Ragged is the default on Blackwell (SM100).
 > On other GPUs, FlashAttention is used as the default.