Drop the remaining customization endpoints that will be replaced by
the dims annotation system:
- LocationInfo dataclass and _get_location_info_of_target_pass_id()
- _comparison_preprocessor() no-op hook
- Simplify main() to use only offset-based step mapping
- Clean up unused imports (re, dataclass, typing)

These imports are only used by main(), and moving them avoids
ModuleNotFoundError when tests import helper functions in environments
without polars installed.

The test_main E2E test calls main() which requires polars for metadata
loading. Skip gracefully in environments without polars.

This reverts commit a05cdef.

Split the monolithic check_tensor_pair() into structured components:

- Add data types: TensorStats, TensorInfo, DiffInfo, TensorComparisonInfo
- Add compare_tensors(): pure computation, returns TensorComparisonInfo
- Add print_comparison(): all printing in one place
- Extract _compute_tensor_stats(), _compute_diff(), _print_diff()
- Update main() to load tensors itself, call compare_tensors + print_comparison
- Remove check_tensor_pair() and _compute_and_print_diff()
- Remove side-effect print from _try_unify_shape()

- Replace test_compute_and_print_diff with test_compute_diff
  (checks DiffInfo fields instead of dict)
- Add test_compute_tensor_stats
- Add test_compare_tensors (new public API)

This reverts commit 296693d.

This reverts commit c71d3a0.

This reverts commit bee3d99.

Extract utility functions from dump_comparator.py into the new
comparator package, removing the _ prefix:
- argmax_coord, compute_smaller_dtype, try_unify_shape,
  calc_rel_diff, load_object

TensorStats, TensorInfo, DiffInfo, TensorComparisonInfo for
separating computation results from printing logic.

compare_tensors returns TensorComparisonInfo with no side effects.
Internal helpers: _compute_tensor_stats, _compute_diff.

printer.py: print_comparison consumes TensorComparisonInfo,
with helpers _print_stats_comparison and _print_diff.
__init__.py re-exports compare_tensors and print_comparison.

entrypoint.py: main() orchestrates metadata reading, row matching,
tensor loading, and calls compare_tensors + print_comparison.
__main__.py: CLI entry point with argparse.

Move 6 existing tests with updated imports to comparator package.
Add new UTs: test_compute_tensor_stats (incl. large tensor),
test_compute_diff (identical + known offset), test_compare_tensors
(normal, shape mismatch, dtype mismatch), test_print_comparison.

- comparator/test_utils.py: utils function tests
- comparator/tensor_comparison/test_compare.py: compare_tensors,
  _compute_diff, _compute_tensor_stats tests
- comparator/tensor_comparison/test_printer.py: print_comparison tests
- comparator/test_entrypoint.py: E2E test with main()

- entrypoint.py: split into main() (CLI) + run(args) (programmatic),
  _parse_args() holds argument definitions, all imports at top level
- __main__.py: thin wrapper calling main()
- test: call run(args) instead of main(args)

- Plain classes instead of unittest.TestCase/CustomTestCase
- assert + pytest.approx instead of self.assertEqual/assertAlmostEqual
- tmp_path fixture instead of tempfile.TemporaryDirectory
- Global imports at top level
- sys.exit(pytest.main([__file__])) entry point

Construct TensorComparisonInfo with deterministic values and assert
exact output for normal case. Assert key substrings for shape
mismatch, downcast, shape unification, and sample printing.

- test_printer: all 5 tests now use full snapshot assertions (== exact
  string), add test_none_quantiles for skipped quantile branch
- test_entrypoint: add capsys assertions (Check count, Skip, rel_diff),
  add filter/no-baseline/step-range scenarios, fix exp_name bug
- test_compare: add quantile value checks, rel_diff known value,
  shape unification, sample generation/suppression tests
- test_utils: add 1D/2D argmax, same-shape noop, trailing mismatch,
  dict-with-value loading, similar/negated tensor rel_diff, unknown dtype

The step offset logic (baseline_step = step - start_id + baseline_start_id)
was confusing and unused in practice. Now baseline_step simply equals the
target step, which is the common case. A clearer --step-offset can be
added later if needed.

Extract _make_tensor_info in test_printer to eliminate 12 repetitive
TensorInfo constructions, and _make_dumper in test_entrypoint to
centralize DumperConfig boilerplate used in 3 places.

- compare.py: extract _quantile_or_none(), remove functools.partial
  and 4 repeated quantile blocks
- printer.py: derive stat_names from dataclasses.fields(TensorStats)
  instead of hardcoded list
- test_entrypoint.py: add num_steps param to _create_dumps, rewrite
  test_step_range to reuse it
- test_printer.py: add comment explaining intentional snapshot
  string duplication

TensorStats, TensorInfo, DiffInfo, TensorComparisonInfo are now
immutable. Restructure compare_tensors to construct TensorInfo after
sample values are known, eliminating post-construction mutation.

This reverts commit e00833c.