CI follow-ups: table compile test compares no values; pretrained-model cache gets evicted #2171

Two follow-ups found while profiling the pytorch test job for #2170.

1. The tablecenternet compile test does not compare outputs

test_torch_compiled_models in tests/pytorch/test_models_table_structure_pt.py says "Compare that outputs are close", but only asserts len(out) == len(compiled_out) == 1 and that the cells/num_rows/num_cols keys exist. The input is random noise, and on it the predictor returns no cells (checked locally), so nothing is compared even in principle. A real table image, or a comparison of the raw model outputs at atol=1e-4 like the detection test does, would make it catch a compile regression.

2. The pretrained-model cache is evicted most of the time

doctr-models-${{ github.job }}-v1 is about 5.7 GB, and the repo's cache limit is 10 GB. In 29 of the last 40 pytest-torch runs the restore missed. A PR run that misses saves its own PR-scoped 5.7 GB copy, which pushes main's entry out, and the next run misses again. A miss costs about 50 to 100 s per run. A possible fix is to restore with actions/cache/restore on pull requests and save only on pushes to main, so PR runs reuse main's copy without evicting it.