Fix loading ThunderKitten libraries (#22)

Summary:
As mentioned in #17

Pull Request resolved: #22

Test Plan:
```
(base) ➜  tritonbench git:(xz9/fix-tk-load) python -c "from tritonbench.utils.loader import load_library; load_library('tk/tk_attn_h100_fwd.so')"
(base) ➜  tritonbench git:(xz9/fix-tk-load) echo $?
0
```

Reviewed By: FindHao

Differential Revision: D65147532

Pulled By: xuzhao9

fbshipit-source-id: 669b2aa4db9b0581a2dbf709e9b577dbbe5e670e

Author: xuzhao9

Diff for: tritonbench/operators/flash_attention/operator.py

@@ -99,7 +99,7 @@
         # colfax Flash Attention V2 for Hopper
         torch.ops.load_library("//ai_codesign/gen_ai/cutlass-kernels:fmha_forward_lib")
     else:
-        from userbenchmark.triton.loader import load_library
+        from tritonbench.utils.loader import load_library
 
         load_library("cutlass_kernels/fmha_forward_lib.so")
     colfax_cutlass_fmha = torch.ops.cutlass.fmha_forward

Diff for: tritonbench/utils/loader.py

@@ -5,5 +5,5 @@ def load_library(library_path: str):
     import torch
 
     prefix, _delimiter, so_file = library_path.partition("/")
-    so_full_path = REPO_PATH.joinpath(prefix, ".data", so_file).resolve()
+    so_full_path = REPO_PATH.joinpath("utils", prefix, ".data", so_file).resolve()
     torch.ops.load_library(str(so_full_path))