fix: respect ratio_list sampling in MultiScaleDataSet - #18345

Problem

MultiScaleDataSet does not correctly respect the _index_map
generated by SimpleDataSet when ratio_list contains values below 1.0.

For ds_width=False, sample indices are resolved through
data_idx_order_list instead of _index_map.

For ds_width=True, wh_ratio_sort is built from the full dataset
instead of the ratio-sampled subset.

After reset_data_lines() changes the sampled subset, the cached
width-aware state in MultiScaleSampler can also become stale.

Example

Given two label files:

• Dataset A: 100 samples
• Dataset B: 100 samples
• ratio_list = [0.5, 0.5]

Expected:

• 50 samples from A
• 50 samples from B

Although _index_map contains the correct sampled indices,
MultiScaleDataSet may access samples outside that subset.

For width-separated samples, this can result in accessing
100 samples from A and 0 samples from B.

Fix

• Resolve MultiScaleDataSet indices through _index_map.
• Build width-aware ordering from the current sampled subset.
• Refresh width-aware ordering after ratio resampling.
• Refresh cached MultiScaleSampler batches after dataset reset.
• Add regression tests for both ds_width=False and ds_width=True,
  including epoch reset behavior.

Tests

python -m pytest -q tests/ppocr/test_simple_dataset.py

Result:

3 passed

All modified files also pass pre-commit.