fix(sampling): preserve remainder samples across shards

[RerankerGuo]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

PR information

SwiftSampling._get_dataset calculated one fixed shard size with integer
division. When the dataset size was not divisible by the number of
data_range shards, every shard used that truncated size and the remainder
samples were never processed.

For example, 10 samples split across 3 processes produced shard sizes
[3, 3, 3], silently dropping sample 9.

This PR distributes the remainder across the first shards. The resulting
shards remain contiguous and ordered, differ in size by at most one, and cover
each input exactly once. Datasets smaller than the shard count are also
handled, with empty shards only after all samples have been assigned.

The change is limited to dataset partitioning for data_range; sampling,
resume, and output behavior are unchanged.

Experiment results

Before the fix:

10 samples / 3 shards -> [3, 3, 3], sample 9 missing
FAILED (Runs=1, success=0, failures=1)

After the fix:

10 samples / 3 shards -> [4, 3, 3], complete ordered coverage
 2 samples / 3 shards -> [1, 1, 0], complete ordered coverage

.venv/bin/python tests/run.py --test_dir tests/sample --pattern test_sampling.py
SUCCESS (Runs=1, success=1)

.venv/bin/pre-commit run --all-files
All hooks passed