Compare full pipeline outputs in the modular tester mixin
`test_save_from_pretrained` and `test_components_auto_cpu_offload_inference_consistent`
compared outputs through `output[0, -3:, -3:, -1]`, which only means "a 3x3 corner of the
first image" for a 4-D `(B, H, W, C)` output. Audio `(B, C, samples)` and 5-D video outputs
still index without error, so a non-image pipeline gets a silently degenerate comparison or
has to override the test.
Compare the whole output instead, as `cosmos` already does by hand in its own override and
as `test_float16_inference` does.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
