【创新应用】新增 QReupload：数据重上传变分量子学习模块（可训练 input scaling + Fourier/贫瘠高原诊断）

[freak-jaeuk]

What

Adds pyqpanda_alg/QReupload/, the package's first supervised estimator whose
circuit parameters are trained against a task loss (sklearn-style fit).
QSVM/QSVR fit only a classical SVM over a parameter-free quantum feature
map; QSVD/QAOA are variational but not supervised ML estimators.

• QReuploadRegressor / QReuploadClassifier — data re-uploading models with
  a trainable input-scaling factor and an optional CNOT entangling ring,
  trained with exact adjoint gradients (vqcircuit.VQCircuit + ADJOINT_DIFF).
• fourier_spectrum(model) — recovers which Fourier harmonics a fitted
  1-feature model learned (the model realises a truncated Fourier series; Schuld,
  Sweke & Meyer, PRA 103, 032430, 2021).
• trainability_scan(qubit_range) — barren-plateau diagnostic,
  Var[∂⟨Z⟩/∂θ] vs qubit count.

Why

Gives pyqpanda-algorithm a trainable QML model (not just kernels), plus two
tools that make encoding expressivity and trainability inspectable before
scaling a circuit.

Validation

• test/QAlgBase/Test_QReupload.py — 10 unit tests pass: behaviour (fit,
  band-limited spectrum, entanglement necessity on a non-separable cross-frequency
  target, classifier, barren-plateau decay, parameter count, multi-feature smoke)
  and estimator-contract checks (rejects wrong feature count / length-mismatched /
  non-finite input, get_params/set_params round-trip).
• example/QAlgBase/testeg_qreupload.py — three demos: (A) 1-feature
  regression + learned Fourier-spectrum readout; (B) an entangle-vs-no-entangle
  ablation on y = 2sin(x0) + 1.5cos(x1) + sin(x0+x1) (the cross term makes
  entanglement load-bearing); (C) a barren-plateau scan. Each is shown next to
  a classical Fourier-ridge baseline — no over-claiming: the classical full-Fourier
  model appears only as an oracle handed the exact basis a priori.

Notes (verified on pyqpanda3 0.3.5)

• Data is baked per-sample as a constant (Param(λ)·float(x)); a product of two
  placeholders (Param·Param) yields a zero adjoint gradient on pyqpanda3 0.3.5.
• Input-scaling factors initialise near 1.0 (≈0 collapses the encoding).
• The module needs only numpy + pyqpanda3; the example imports sklearn for
  the classical comparator (same as the existing QSVR/QKmeans examples).
• fourier_spectrum is documented as an empirical DFT: the strict integer-degree
  bound holds for unit input scaling, since a trained scaling factor can move peaks
  off the integer grid.

The new subpackage is registered in pyqpanda_alg/__init__.py alongside the
existing QSVM/QSVD/… modules; the test follows this repo's test/pytest.ini
(Test_*.py under test/QAlgBase/) and Test(feature, content) case naming.