Fix variable-length M4 series loading

[kiwoongyoon]

Title: Fix variable-length M4 series loading

Summary

• Construct filtered M4 series with an object dtype so series of different lengths can be loaded.
• Preserve the existing per-series list behavior without changing sampling or batching.
• Checked the issue snapshot for duplicate work; no related pull request was found.

Test evidence

• Focused Dataset_M4 base/patched regression harness (exit 0): reproduced the base ragged-series failure and verified the patch preserved two unequal-length series.
• Focused AST assertion for the training_values object dtype (exit 0)
• python3 -m compileall -q data_provider/data_loader.py (exit 0)
• git diff --check (exit 0)

Risks or notes for maintainers

• The sandbox did not have the repository's Python dependencies and could not fetch them because network access was unavailable, so the regression harness used standard-library stubs around the actual Dataset_M4 class body.
• This restores the same object-array approach previously used by repository commit 13680a6 for mismatched M4 sample sizes.

Fixes #850