don't replace TensorImpl for inplace min/max dim (#35591)
Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/35591
Test Plan: buck test mode/dev //caffe2/test:cuda -- 'test_dim_reduction_cpu \(test_torch\.TestTorchDeviceTypeCPU\)'
Differential Revision: D20718321
Pulled By: ngimel
fbshipit-source-id: eba09b37ab1f22463114da69d35982e881dcaa85
