[docs] feat: improve docstrings in tensordict_utils.py (#1345) (#4732)

## Summary
- Add comprehensive Google-style docstrings to 17 functions in
`verl/utils/tensordict_utils.py`
- Improves IDE support with parameter hints, type information, and
examples
- Part of the documentation standardization effort (Issue #1345)

## Functions Documented
**Non-tensor data handling:**
- `assign_non_tensor_data` - Assign single non-tensor value to
TensorDict
- `unwrap_non_tensor_data` - Unwrap NonTensorData to get underlying
value
- `get_non_tensor_data` - Retrieve and unwrap non-tensor data

**TensorDict concatenation/splitting:**
- `concat_nested_tensors` - Concatenate 2D nested tensors
- `concat_tensordict_with_none_bsz` - Handle TensorDicts with empty
batch size
- `concat_tensordict` - Concatenate multiple TensorDicts along dim 0
- `chunk_tensordict` - Split TensorDict into equal-sized chunks

**Data access and manipulation:**
- `index_select_tensor_dict` - Select rows using indices
- `union_tensor_dict` - Merge two TensorDicts
- `make_iterator` - Create mini-batch iterator for training
- `assert_tensordict_eq` - Assert two TensorDicts are equal
- `get` / `get_keys` - Get values with automatic unwrapping
- `pop` / `pop_keys` - Remove and return values

**Padding utilities:**
- `pad_to_divisor` - Pad batch dimension for distributed training
- `unpad` - Remove padding from TensorDict

## Test plan
- [ ] Existing tests pass (documentation-only changes)
- [ ] Docstrings follow Google-style format consistent with other files

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-authored-by: yurekami <[email protected]>
Co-authored-by: Claude Opus 4.5 <[email protected]>

Author: yurekami