[Code scan] Handle small workloads in MPI range splitting

This issue comes from a Codex global scan of deepmodeling/tbplas at commit 4d3652b3dba8eb930e4869a6d41074f62d703692.

Severity: Medium

`dist_range()` delegates to `split_range(n_max, num_group=self.__size)`. The split helper asserts that `num_group` is no larger than the number of elements. As a result, MPI runs with more ranks than work items can fail before callers have a chance to skip empty partitions. `dist_bound()` also assumes the assigned range is non-empty and calls `min()`/`max()` on it.

Code references:
https://github.com/deepmodeling/tbplas/blob/4d3652b3dba8eb930e4869a6d41074f62d703692/tbplas/parallel.py#L107-L120
https://github.com/deepmodeling/tbplas/blob/4d3652b3dba8eb930e4869a6d41074f62d703692/tbplas/utils.py#L333-L340
https://github.com/deepmodeling/tbplas/blob/4d3652b3dba8eb930e4869a6d41074f62d703692/tbplas/utils.py#L313-L329

Suggested fix: allow empty ranges when `num_group > n_max`, and make `dist_bound()` return a documented empty sentinel or require callers to use `dist_range()` when empty partitions are possible.

Start with dist_range() in tbplas/parallel.py and trace its split_range() call in tbplas/utils.py. Inspect dist_bound() in the same file, then reproduce an MPI run where the number of ranks exceeds the work items. Done means empty partitions no longer fail, and dist_bound() has a documented behavior or callers use dist_range() when empties are possible.

Written by the indexing model from the issue text.