[Code scan] LMDBDataset loses the source LMDB path when multiple roots match

This issue is part of a Codex global repository scan.

Problem:
`LMDBDataset.__init__` counts entries per matched LMDB path, but records only the logical `file` and local index. `get()` recomputes all matching LMDB paths for that file and loops through all of them, overwriting the loaded sample each time.

Code references:
https://github.com/deepmodeling/deeptb/blob/86c60c73996f0dd961c3138f2e88424382cb734e/dptb/data/dataset/lmdb_dataset.py#L69-L77
https://github.com/deepmodeling/deeptb/blob/86c60c73996f0dd961c3138f2e88424382cb734e/dptb/data/dataset/lmdb_dataset.py#L138-L144
https://github.com/deepmodeling/deeptb/blob/86c60c73996f0dd961c3138f2e88424382cb734e/dptb/data/dataset/lmdb_dataset.py#L180-L187

Impact:
For wildcard or multi-root datasets where the same LMDB folder name appears under multiple roots, samples from earlier databases can be read from the last matching LMDB instead. This corrupts dataset indexing and can mix labels/geometries from the wrong source.

Suggested fix:
Record the resolved LMDB path alongside each global sample index, and have `get()` open exactly that path instead of recomputing all matches.

Start in dptb/data/dataset/lmdb_dataset.py at the __init__ and get() sections referenced in the issue, and trace how matched LMDB paths map to global sample indices. Verify that each indexed sample retains its resolved LMDB path and that get() reads only from that path, preventing samples from one database being replaced by another.

Written by the indexing model from the issue text.