fix: keep TCN TS singleton batch outputs 1d

[he-yufeng]

Summary

• Keep the time-series TCN model output as a 1-D batch vector for singleton batches.
• Add a regression test that checks batch size 1 and batch size 2 both keep the batch dimension.

Fixes #1752.

To verify

• python -m ruff check qlib/contrib/model/pytorch_tcn_ts.py tests/model/test_pytorch_tcn_ts.py
• python -m py_compile qlib/contrib/model/pytorch_tcn_ts.py tests/model/test_pytorch_tcn_ts.py

Notes

• python -m pytest tests/model/test_pytorch_tcn_ts.py -q is blocked in my Windows checkout because qlib requires its Cython extensions to be compiled before import. python -m pip install -e . --no-deps reaches the qlib.data._libs.rolling build, but MSVC fails while reading the Python headers with missing io.h.

[he-yufeng]

Quick follow-up: this is still scoped to the singleton-batch TCN TS output shape, with a focused regression test in tests/model/test_pytorch_tcn_ts.py.

Could a maintainer take a look when convenient?