fix(model): group DailyBatchSampler batches by datetime positions

[张弘 (shuxue6662-a11y)]

Description

DailyBatchSampler assumed that the rows of a trading day are contiguous in the data source: it computed per-day counts via groupby("datetime").size(), converted them to start offsets with cumsum, and sliced contiguous ranges np.arange(idx, idx + count).

TSDataSampler.get_index() swaps the level labels back to (datetime, instrument) but does NOT reorder the rows, which remain instrument-major (see "Indices design" in TSDataSampler). As a result each "daily" batch actually contained one instrument across many consecutive days instead of one day's cross-section.

This PR changes DailyBatchSampler to group the actual row positions by the datetime level, so every batch contains exactly one trading day's cross-section. It also:

• fixes GATs.predict() to align predictions with the sampler's iteration order (dl_test.get_index()[np.concatenate(sampler_test.batches)]), since the previous dl_test.get_index() alignment assumed the old row order;
• fixes __len__ to return the number of batches instead of the number of samples;
• clarifies the TSDataSampler.get_index() docstring about the instrument-major row order;
• adds regression tests covering the one-day-per-batch invariant, full sample coverage, and batch count.

Motivation and Context

Fixes #2319

How Has This Been Tested?

• New regression tests: python -m pytest tests/test_daily_batch_sampler.py -> 2 passed
• black --check -l 120 on the changed files
• flake8 on the changed files
• Full pipeline test (pytest qlib/tests/test_all_pipeline.py) -- requires downloading test data; covered by CI

Types of changes

• Fix bugs
• Add new feature
• Update documentation

[张弘 (shuxue6662-a11y)]

@microsoft-github-policy-service agree