[1.2] Batched torch component library (1D + 2D)

[CSSFrancis]

Batched, autograd-differentiable ports of the hyperspy components. Each takes (x, params) → (P, C) fully batched, no Python loop over items.

1D (hyperspy.components1d): Gaussian, GaussianHF, Lorentzian, Voigt, SplitVoigt, PowerLaw, Offset, Polynomial, Exponential, Arctan, Erf, Doniach, SkewNormal, Logistic, HeavisideStep, ScalableFixedPattern.

2D (hyperspy.components2d): Gaussian2D, Expression.

Every component gets a parity test against the hyperspy component it mirrors (same params → same curve within tolerance). Mark which are linear in amplitude — [1.3]'s variable-projection path keys off that.

Part of #50.

---

Plan: RELEASE_0_3_0_PLAN.md

[CSSFrancis]

Delivered in 0.3.0, merged to main via #89.

72e2f0c — feat(fitting): batched GPU model fitting (0.3.0 Wave 1.1-1.3)

Closing from the merge rather than automatically: the commits cite issues as (#52) rather than Closes #52, so GitHub linked none of them (#89 reported 0 closing references).