[rope] Fix running with cudagraph

[xuzhao9]

Tensor view does not support in-place detaching with .detach_(), use .contiguous() call to convert tensor views to tensor.

Fixes #155

Test plan:

$ python run.py --op rope  --cudagraph
       (H, T)    apply_rotary_pos_emb-latency    liger_rotary_pos_emb-latency    inductor_rotary_pos_emb_full_op-latency
-------------  ------------------------------  ------------------------------  -----------------------------------------
 (8192, 1024)               0.209794 (±0.08%)               0.079294 (±0.09%)                          0.044696 (±0.17%)
 (8192, 2048)               0.425296 (±0.07%)               0.167398 (±0.07%)                          0.083712 (±0.10%)
 (8192, 4096)               0.869108 (±0.07%)               0.333221 (±0.04%)                          0.175674 (±0.27%)
 (8192, 8192)               1.734228 (±0.05%)               0.655273 (±0.05%)                          0.442072 (±1.22%)
(8192, 16384)              4.405540 (±13.46%)               1.325537 (±8.78%)                          1.177186 (±0.31%)
  (512, 2048)               0.033462 (±0.33%)               0.010369 (±0.14%)                          0.006766 (±0.16%)
 (2048, 2048)               0.098888 (±0.23%)               0.027084 (±0.25%)                          0.019540 (±1.82%)
 (8192, 2048)               0.424953 (±0.06%)               0.167613 (±0.05%)                          0.088726 (±0.66%)

[facebook-github-bot]

@xuzhao9 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

[facebook-github-bot]

@xuzhao9 merged this pull request in a42aa62.