benchmark/tutorial lazy resampling components and workflows

[wyli]

Describe the bug
would be great to provide some measurements about lazy resampling output quality, speed, and memory footprint on cpu/gpu.

• basic comparisons for fusing a few spatial transforms (adds a lazy resampling functional example tutorials#1254)
• testing with pytorch/monai dataloaders (https://gist.github.com/wyli/2e1d9d22e4989f4a055dd4f256a02ea9/c3a2ddae3dc73f8780f53df26659d014c024b859 update the compose example lazy resampling tutorials#1321)
• benchmark with typical segmentation workflows (Add lazy-resampling benchmark tutorial tutorials#1322)
• benchmark with model zoo workflows (Update wholeBody_ct_segmentation model-zoo#397)

please note that the Compose/caching enhancements are still WIP, the logic may need to be tuned accordingly.

cc @yiheng-wang-nv @KumoLiu @Nic-Ma