Fix daily batching for time-series samplers

[Serhii Ovsiienko (sergio12S)]

TSDataSampler.get_index() exposes datetime/instrument labels over an
instrument-major row layout. DailyBatchSampler previously calculated correct
per-date counts but converted them to contiguous ranges, so a nominal daily
batch could mix multiple dates.

I reproduced this with five dates and three instruments: every emitted batch
contained three distinct dates. This change groups the actual source-row
positions by datetime. Because that changes iteration order, predict() now
builds its result index from the sampler's order as well, preventing otherwise
correct predictions from being assigned to the wrong observations. __len__
now returns the number of daily batches.

The added regression test constructs a real TSDataSampler and checks batch
membership, sampler length, and output-index alignment.

Tested with:

python -m pytest -q tests/model/test_pytorch_gats_ts.py — 1 passed.

Closes #2319.

[Parafee41 (koriyoshi2041)]

Verified at 1d58db7. Grouping source-row positions by datetime fixes the instrument-major layout bug, and using the same sampler order for the prediction index keeps values aligned. The regression covers daily membership, batch count, and index ordering; the focused test passes locally, and the diff check is clean.