fix: transform/convert dead ends and JSON round-trip (#537)

[Shashankss1205]

Fixes #537. Four transform/convert dead ends:

BUG-21 — np.ndarray handles are a dead end

Converting a handle to an index-less mtype (np.ndarray, numpy3D, ...) produced a handle with no time index that broke inspect_data/split_data/format and fabricated a cutoff. Those targets are now rejected up front with a message pointing to pd.Series/pd.DataFrame.

NB-09 — Series→Panel convert dumped a raw mtype error

convert(Series handle -> pd-multiindex) returned sktime's multi-paragraph "No valid mtype could be identified ..." dump. Caught and replaced with a clean domain message: "Cannot convert a Series-scitype handle to 'pd-multiindex'. That target expects a different scitype (Panel)."

NB-10 — JSON was save-only

• Added a JSON reader to the file adapter (.json → json, pd.read_json records orient); the load error hint no longer points into a dead end.
• save_data json now writes the index as a "time" column (records orient dropped it), so a saved JSON round-trips through load_data_source(time_column="time").

BUG-19 — duplicate timestamps hard-rejected before auto-format

Duplicate timestamps were a validation error, so no handle was returned and the documented remedy (transform_data(action='format', remove_duplicates=True)) was unreachable. Downgraded to a warning, so the handle loads and auto-format de-duplicates it (keeping the first of each).

Testing

• New tests/test_transform_deadends.py (5 tests): ndarray target rejected; Series→Panel clean error (no mtype dump); valid convert works; duplicate-timestamp load succeeds and is de-duplicated; JSON save→load round-trip.
• Full suite: 285 passed.

🤖 Generated with Claude Code

[Shashankss1205]

Closing: this PR's content is already on main as commit caaa83b (identical patch-id; it was landed directly rather than through the PR merge button). Nothing further to merge.