Fix implicit fallthrough warnings in FeatureLPPooling.cu (#25451)
Summary:
`-Wimplicit-fallthrough` is enabled for recent GCC versions, and there's about 1000 lines of warnings in the build output with GCC 9.1 like:
```
/home/rgommers/code/pytorch/aten/src/THCUNN/FeatureLPPooling.cu: In function ‘bool runFeatureLPPoolingUpdateOutput(THCState*, const THCDeviceTensor<T, 4>&, THCDeviceTensor<T, 4>&, float, int, int) [with T = c10::Half]’:
/home/rgommers/code/pytorch/aten/src/THCUNN/FeatureLPPooling.cu:474:10: warning: this statement may fall through [-Wimplicit-fallthrough=]
474 | L2_WIDTH_CASE(2);
| ^~~~~~
/home/rgommers/code/pytorch/aten/src/THCUNN/FeatureLPPooling.cu:475:1: note: here
475 | L2_WIDTH_CASE(3);
| ^
...
/home/rgommers/code/pytorch/aten/src/THCUNN/FeatureLPPooling.cu:639:11: warning: this statement may fall through [-Wimplicit-fallthrough=]
639 | LP_WIDTH_CASE(15);
| ^~~~~~
/home/rgommers/code/pytorch/aten/src/THCUNN/FeatureLPPooling.cu:640:1: note: here
640 | LP_WIDTH_CASE(16);
| ^
```
Fix by ending each case statement with `break;`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25451
Differential Revision: D17131254
Pulled By: ezyang
fbshipit-source-id: 55b513620438cbbf86052f22d799d790b0633fa2
