fix(v1): respect batch_size in train_gen when permute=False

[staticpayload]

Summary

• fix TimeSeriesdata.train_gen() in v1/src/timesfm/data_loader.py so permute=False yields deterministic batch windows instead of all series every step
• keep existing permute=True sampling behavior unchanged
• add regression tests that validate index windows, tensor leading dimensions, and no extra batch on even splits

Bug

Issue #359 reports that train_gen ignores batch_size when permute=False by setting tsidx = np.arange(num_ts) every iteration.

This makes non-permuted training load all series in each generated sample, which breaks expected batching semantics and inflates memory/compute.

Fix

For permute=False, iterate explicit batch starts (range(0, num_ts, self.batch_size)) and slice tsidx as:
np.arange(batch_idx, min(batch_idx + self.batch_size, num_ts)).

Tests

Added v1/tests/test_data_loader.py with two regression tests:

1. test_train_gen_non_permute_respects_batch_windows
2. test_train_gen_non_permute_no_extra_batches_on_even_split

Local run:

python3 -m pytest -q v1/tests/test_data_loader.py
# 2 passed

Fixes #359