feat(pt/dpmodel): add lmdb dataloader (#5283)

<!-- This is an auto-generated comment: release notes by coderabbit.ai
-->
## Summary by CodeRabbit

* **New Features**
* Full LMDB data support across training, validation and testing with
PyTorch-friendly dataset/dataloader integration, per-atom-count (nloc)
grouping, and deterministic distributed batching.

* **Tests**
* Extensive unit and integration tests covering LMDB reader,
dataset/dataloader, samplers, collation, and mixed-/uniform-nloc
scenarios.

* **Documentation**
* Added an example LMDB training configuration and updated test runner
to handle LMDB-grouped test runs.

* **Chores**
  * Added runtime dependencies for LMDB and msgpack support.
<!-- end of auto-generated comment: release notes by coderabbit.ai -->

Author: iProzd

deepmd/dpmodel/utils/__init__.py

@@ -6,6 +6,14 @@
     AtomExcludeMask,
     PairExcludeMask,
 )
+from .lmdb_data import (
+    DistributedSameNlocBatchSampler,
+    LmdbDataReader,
+    LmdbTestData,
+    SameNlocBatchSampler,
+    is_lmdb,
+    make_neighbor_stat_data,
+)
 from .network import (
     EmbeddingNet,
     FittingNet,
@@ -44,13 +52,17 @@
 
 __all__ = [
     "AtomExcludeMask",
+    "DistributedSameNlocBatchSampler",
     "EmbeddingNet",
     "EnvMat",
     "FittingNet",
+    "LmdbDataReader",
+    "LmdbTestData",
     "NativeLayer",
     "NativeNet",
     "NetworkCollection",
     "PairExcludeMask",
+    "SameNlocBatchSampler",
     "aggregate",
     "build_multiple_neighbor_list",
     "build_neighbor_list",
@@ -59,10 +71,12 @@
     "get_graph_index",
     "get_multiple_nlist_key",
     "inter2phys",
+    "is_lmdb",
     "load_dp_model",
     "make_embedding_network",
     "make_fitting_network",
     "make_multilayer_network",
+    "make_neighbor_stat_data",
     "nlist_distinguish_types",
     "normalize_coord",
     "phys2inter",