[Python][C++] Fix tree-model TsFileDataFrame returning empty data on reused readers and uppercase names

[Young-Leo]

Summary

When reading tree-model TsFiles through TsFileDataFrame, series reads could
return empty arrays in two situations. This PR fixes both, which together were
responsible for "read one series OK, then read another series returns empty" on
a reused reader.

Both fixes go on top of the existing tree-model support (#816).

---

Fix 1 — C++: make StringArrayDeviceID::split_table_name() idempotent

Problem
Device IDs are cached and reused across queries. split_table_name() /
init_prefix_segments() appended to prefix_segments_ every time it was
called, so on a reused reader the prefix segments accumulated across successive
queries. The stale/duplicated prefix segments corrupted the per-device path
columns, so a second series read on the same reader could match nothing and
return empty (and it also leaked the previously allocated segment pointers).

Change
init_prefix_segments() now clears and frees the existing prefix_segments_
before rebuilding, making split_table_name() idempotent. Repeated calls
produce a stable, correct segment vector.

Files: cpp/src/common/device_id.cc, test cpp/test/common/device_id_test.cc
(DeviceIdTest.SplitTableNameIsIdempotent).

---

Fix 2 — Python: read series with uppercase measurement / path names

Problem
Tree TsFiles have no table schema, so the Python dataset layer reads them by
synthesizing a virtual table via query_table_on_tree. Because the table model
is case-insensitive, the native layer normalizes the synthesized column names
and path segments to lower case (e.g. Temperature -> temperature), while
the dataset catalog keeps the original casing from timeseries metadata. The two
tree read paths then looked the field column and device path up
case-sensitively, so an uppercase measurement or path segment failed the lookup
and returned an empty array:

with TsFileDataFrame(path) as tsdf:
    tsdf["root.ln.wf01.wt01.Temperature"][:]
    # -> array([], dtype=float64)   # expected [0.5, 1.5, 2.5, 3.5, 4.5]

Change Match field column names and device path segments case-insensitively in _read_series_by_row_tree and _read_arrow_tree, consistent with the case-insensitive table model. The original casing is still surfaced to users; only the internal lookup against the lower-cased result set is normalized.

Files: python/tsfile/dataset/reader.py, regression test in python/tests/test_tsfile_dataset.py.

---

Tests

• New C++ regression test: DeviceIdTest.SplitTableNameIsIdempotent.
• New Python regression tests covering reused-reader reads and uppercase measurement/path names.
• No public API change; logical series names keep their original casing.

[ColinLeeo]

It's not a good idea but useful...

[ColinLeeo]

In the tree model, both device paths and measurements are case-sensitive, so we should not simply normalize them to lowercase. Otherwise, case-distinct paths/measurements may be overwritten or mixed. The Python layer should preserve the original casing and resolve values by exact column position. On the C++ side, the tree query path does not seem to lowercase these names directly, but ResultSet name-based lookup is case-insensitive, so there may still be a similar collision risk that we should verify/fix as well.