Inaccurate API Docstrings for Attention Prefill

[bkryu]

Some descriptions provided in docstrings and API documentations are incorrect:

1. BatchPrefillWithPagedKVCacheWrapper states "The implementation backend, could be auto/fa2,fa3 or cudnn.", but in fact allows trtllm-gen as a backend.
   • However, the plan function of BatchPrefillWithPagedKVCacheWrapper returns an incorrect output when causal=False. It is suspected that trtllm-gen backend runs causal=True. Example from flashinfer_benchmark.py below:

## Causal case:
$ python3 flashinfer_benchmark.py --routine BatchPrefillWithPagedKVCacheWrapper --backends fa2 cudnn trtllm-gen --page_size 16 --batch_size 1 --s_qo 8192 --s_kv 8192 --num_qo_heads 64 --num_kv_heads 8 --head_dim_qk 128 --head_dim_vo 128 --random_actual_seq_len --refcheck --q_dtype bfloat16 --kv_dtype bfloat16 --causal --allow_output_mismatch --generate_repro_command
[INFO] To reproduce this test case, run the following command: python3 flashinfer_benchmark.py --routine BatchPrefillWithPagedKVCacheWrapper --backends fa2 cudnn trtllm-gen --page_size 16 --batch_size 1 --s_qo 8192 --s_kv 8192 --num_qo_heads 64 --num_kv_heads 8 --head_dim_qk 128 --head_dim_vo 128 --random_actual_seq_len --refcheck --q_dtype bfloat16 --kv_dtype bfloat16 --causal --allow_output_mismatch --generate_repro_command
[PERF] fa2            :: median time 2.492 ms; std 0.010 ms; achieved tflops 347.593 TFLOPs/sec; achieved tb_per_sec 0.108 TB/sec
[PERF] cudnn          :: median time 0.862 ms; std 0.019 ms; achieved tflops 1005.157 TFLOPs/sec; achieved tb_per_sec 0.311 TB/sec
[PERF] trtllm-gen     :: median time 0.841 ms; std 0.027 ms; achieved tflops 1030.423 TFLOPs/sec; achieved tb_per_sec 0.319 TB/sec

## Same dimensions but non-causal case
# Output mismatch happens and time is nearly identical to causal case when exepected to be 2x
python3 flashinfer_benchmark.py --routine BatchPrefillWithPagedKVCacheWrapper --backends fa2 cudnn trtllm-gen --page_size 16 --batch_size 1 --s_qo 8192 --s_kv 8192 --num_qo_heads 64 --num_kv_heads 8 --head_dim_qk 128 --head_dim_vo 128 --random_actual_seq_len --refcheck --q_dtype bfloat16 --kv_dtype bfloat16 --allow_output_mismatch --generate_repro_command
[INFO] To reproduce this test case, run the following command: python3 flashinfer_benchmark.py --routine BatchPrefillWithPagedKVCacheWrapper --backends fa2 cudnn trtllm-gen --page_size 16 --batch_size 1 --s_qo 8192 --s_kv 8192 --num_qo_heads 64 --num_kv_heads 8 --head_dim_qk 128 --head_dim_vo 128 --random_actual_seq_len --refcheck --q_dtype bfloat16 --kv_dtype bfloat16 --allow_output_mismatch --generate_repro_command
[ERROR] Output tensor mismatch between backends fa2 and trtllm-gen: 26629186 / 59564032 (44.71%) elements are different
[PERF] fa2            :: median time 4.684 ms; std 0.005 ms; achieved tflops 369.869 TFLOPs/sec; achieved tb_per_sec 0.057 TB/sec
[PERF] cudnn          :: median time 1.442 ms; std 0.039 ms; achieved tflops 1201.270 TFLOPs/sec; achieved tb_per_sec 0.186 TB/sec
[PERF] trtllm-gen     :: median time 0.849 ms; std 0.033 ms; achieved tflops 2041.414 TFLOPs/sec; achieved tb_per_sec 0.316 TB/sec

2. BatchPrefillWithRaggedKVCacheWrapper states that backends auto, fa2, fa3 and trtllm-gen are supported, but:
   • cutlass is also supported but omitted.
   • trtllm-gen is not supported an outputs it is not supported yet (message printed here)