Updated regerssion sampler

[AbasKhan]

This PR introduces a configurable uniform split sampler for JSONL datasets, along with its configuration and tests. 🎯

---

Adds UniformSplitSampler, which:

• 📥 Ingests one or more input JSONL files.
• 🔢 Requires a numeric integer score field per example and normalizes these scores.
• ✂️ Produces balanced train/validation JSONL outputs according to the configured split ratio.

---

_build_splits

• 🧩 Groups examples by label and computes per-label quotas for train/validation.
• 🧷 Uses split_label_pools to partition each label’s pool into train/val.
• 📈 Applies sampling with a max_oversampling_ratio cap so rare labels are upsampled but not duplicated arbitrarily.
• 🔀 Shuffles the resulting splits to avoid ordering artifacts.
• 📊 Logs overall and per-label distributions for transparency and easier debugging.

---

In short: this sampler helps you build balanced, reproducible, and well-logged train/val splits from JSONL.

[ajude2s]

Redundant. There is a check above if pool.empty or target <= 0.
max_allowed will be less than or equal to 0 only if max_oversampling_ratio is less than 0.

Which would never be the case, right?

[ajude2s]

Awesome work, Abbas. 👍
I have added minor changes and some suggestions.

Also, should we not add the original sampler (the fixed distribution which yields "best" performance) as well?