Add stacked layers, dropout, benchmark, and contributor docs

[TravisCao]

Summary

Extends the qlstm library toward torch.nn.LSTM parity and hardens the repo
for contributors, without breaking the existing single-layer API.

Changes

Stacking and regularisation

• num_layers on QLSTM/LQLSTM stacks recurrent layers. Each layer reads the
  hidden-state sequence of the layer below; output is the top layer's sequence
  and h_n/c_n are shaped (num_layers, batch, hidden_size), as in
  torch.nn.LSTM.
• dropout applies to the output of every layer except the last.
• The single-layer path and the layer.cell attribute are unchanged, so
  existing code keeps working.

Tests

• New coverage for stacked shapes, gradient flow to every layer, stacked initial
  state, the wrong-layer-count error, eval-time dropout determinism, the cell
  property, and constructor validation. Full suite: 18 passed.

Examples

• examples/benchmark_vs_classical.py runs a fair side-by-side against
  torch.nn.LSTM on the same running-sum task and prints parameter counts and
  train/validation loss. It is a reproducible comparison, not a performance
  claim.

Docs and metadata

• README documents stacking and links the benchmark; CHANGELOG records the
  changes.
• Community health files: CONTRIBUTING.md, issue templates, pull-request
  template.
• Python 3.13 packaging classifier, which CI already tests.

Test plan

• pytest -q green locally (18 passed).
• CI runs the same suite on Python 3.10 through 3.13.