Package the quantum LSTM as an installable, tested library

[TravisCao]

This turns the repository into a reusable open-source library while keeping the original paper-reproduction code intact.

What changed

• New top-level qlstm package: QLSTM, LQLSTM, QLSTMCell, and a make_vqc factory. The layers mirror the torch.nn.LSTM interface, so they drop into existing models.
• Built on current PyTorch and PennyLane (tested with torch 2.13, PennyLane 0.45). Each LSTM gate is an encode-entangle-measure variational circuit built from PennyLane templates, which keeps it batching-safe and differentiable; gradients reach both the circuit weights and the inputs.
• MIT LICENSE, pyproject.toml (so pip install qlstm works), and a machine-readable CITATION.cff.
• Test suite (11 passing): output shapes, gradient flow to inputs and every parameter, a learning check, and the API surface.
• Runnable examples/quickstart.py, CHANGELOG.md, and CI on Python 3.10-3.12.
• README rewritten around the library; the dataset notes are folded into a details block. The paper reproduction stays under src/.
• The dead root __init__.py (a relative import that could never resolve, and which broke test collection) is now empty.

Verification

pytest -q → 11 passed. python examples/quickstart.py drives the loss from 0.10 to 0.0006. uv build produces a valid wheel and sdist.