benchmark: include convert latency in bench_append_paged_kv_cache (#590)

```
model: l1b      seqlens: [1, 1, 1, 1, 1, 1, 1, 1]                 convert: 45us 1layer:  7us 16layers: 151us throughput:    4.936GB/s
model: l1b      seqlens: [4993, 1, 1, 1, 1, 1, 1, 1]              convert: 42us 1layer: 14us 16layers: 271us throughput: 1434.769GB/s
model: l1b      seqlens: [5000]                                   convert: 44us 1layer: 14us 16layers: 272us throughput: 1438.581GB/s
model: l1b      seqlens: [625, 625, 625, 625, 625, 625, 625, 625] convert: 46us 1layer: 14us 16layers: 274us throughput: 1440.357GB/s
---
model: l3b      seqlens: [1, 1, 1, 1, 1, 1, 1, 1]                 convert: 42us 1layer:  7us 28layers: 226us throughput:    9.946GB/s
model: l3b      seqlens: [4993, 1, 1, 1, 1, 1, 1, 1]              convert: 43us 1layer: 22us 28layers: 647us throughput: 1896.687GB/s
model: l3b      seqlens: [5000]                                   convert: 42us 1layer: 22us 28layers: 646us throughput: 1898.796GB/s
model: l3b      seqlens: [625, 625, 625, 625, 625, 625, 625, 625] convert: 41us 1layer: 22us 28layers: 648us throughput: 1890.115GB/s
---
model: l8b      seqlens: [1, 1, 1, 1, 1, 1, 1, 1]                 convert: 41us 1layer:  7us 32layers: 252us throughput:    9.940GB/s
model: l8b      seqlens: [4993, 1, 1, 1, 1, 1, 1, 1]              convert: 42us 1layer: 21us 32layers: 730us throughput: 1905.826GB/s
model: l8b      seqlens: [5000]                                   convert: 41us 1layer: 22us 32layers: 729us throughput: 1903.697GB/s
model: l8b      seqlens: [625, 625, 625, 625, 625, 625, 625, 625] convert: 47us 1layer: 22us 32layers: 737us throughput: 1899.630GB/s
---
model: l70b-tp8 seqlens: [1, 1, 1, 1, 1, 1, 1, 1]                 convert: 42us 1layer:  6us 80layers: 552us throughput:    1.283GB/s
model: l70b-tp8 seqlens: [4993, 1, 1, 1, 1, 1, 1, 1]              convert: 41us 1layer:  9us 80layers: 800us throughput:  539.484GB/s
model: l70b-tp8 seqlens: [5000]                                   convert: 41us 1layer:  9us 80layers: 788us throughput:  548.648GB/s
model: l70b-tp8 seqlens: [625, 625, 625, 625, 625, 625, 625, 625] convert: 41us 1layer: 10us 80layers: 803us throughput:  537.731GB/s
```

benchmarks/bench_append_paged_kv_cache.py

@@ -1,6 +1,6 @@
 import argparse
 import dataclasses
-from typing import cast
+from typing import Tuple, cast
 
 import torch
 from triton.testing import do_bench
@@ -99,14 +99,19 @@ def main():
                 dtype=torch.int32,
             )
 
-            batch_indices, positions = flashinfer.get_batch_indices_positions(
-                x_indptr,
-                flashinfer.get_seq_lens(kv_indptr, kv_last_page_len, page_len),
-                k.shape[0],
-            )
+            @torch.cuda.nvtx.range(f"convert model={model_name}, seqlens={seqlens}")
+            def fn_convert() -> Tuple[torch.Tensor, torch.Tensor]:
+                return flashinfer.get_batch_indices_positions(
+                    x_indptr,
+                    flashinfer.get_seq_lens(kv_indptr, kv_last_page_len, page_len),
+                    k.shape[0],
+                )
+
+            batch_indices, positions = fn_convert()
+            convert_latency_ms = cast(float, do_bench(fn_convert))
 
-            @torch.cuda.nvtx.range(f"model={model_name}, seqlens={seqlens}")
-            def fn():
+            @torch.cuda.nvtx.range(f"append model={model_name}, seqlens={seqlens}")
+            def fn() -> None:
                 flashinfer.append_paged_kv_cache(
                     k,
                     v,
@@ -120,7 +125,7 @@ def fn():
                 )
 
             latency_ms = cast(float, do_bench(fn))
-            all_layers_latency_ms = latency_ms * model.num_layers
+            all_layers_latency_ms = convert_latency_ms + latency_ms * model.num_layers
             throughput = (
                 k.numel()
                 * k.element_size()
@@ -131,8 +136,9 @@ def fn():
             print(
                 f"model: {model_name:8}",
                 f"seqlens: {seqlens!r:{seqlen_strlen}}",
-                f"single_layer: {latency_ms:5.3f}ms",
-                f"all_layers: {all_layers_latency_ms:7.3f}ms",
+                f"convert: {convert_latency_ms*1e3:2.0f}us",
+                f"1layer: {latency_ms*1e3:2.0f}us",
+                f"{model.num_layers}layers: {all_layers_latency_ms*1e3:3.0f}us",
                 f"throughput: {throughput*1e-9:8.3f}GB/s",
             )
         print("---")