docs: add formulas to fitting classes

Add mathematical formulas to fitting class docstrings:
- DipoleFitting: dipole computation from rotation matrix
- PolarFitting: polarizability tensor with diagonal/full fitting

Follow numpydoc convention: parameters in class docstring.

Author: njzjz-bot

deepmd/dpmodel/fitting/dipole_fitting.py

@@ -34,7 +34,23 @@
 @BaseFitting.register("dipole")
 @fitting_check_output
 class DipoleFitting(GeneralFitting):
-    r"""Fitting rotationally equivariant diploe of the system.
+    r"""Fitting rotationally equivariant dipole of the system.
+
+    The dipole :math:`\boldsymbol{\mu}` is computed from the fitting network output
+    and the rotation matrix:
+
+    .. math::
+        \boldsymbol{\mu}^i = \mathbf{M}^i \cdot \mathbf{R}^i,
+
+    where :math:`\mathbf{M}^i \in \mathbb{R}^{1 \times m_1}` is the output of the fitting
+    network for atom :math:`i`, and :math:`\mathbf{R}^i \in \mathbb{R}^{m_1 \times 3}` is
+    the rotation matrix from the descriptor. The fitting network is:
+
+    .. math::
+        \mathbf{M}^i = \mathcal{L}^{(n)} \circ \mathcal{L}^{(n-1)} \circ \cdots \circ \mathcal{L}^{(0)}(\mathcal{D}^i),
+
+    where :math:`\mathcal{D}^i` is the descriptor and each layer :math:`\mathcal{L}^{(k)}`
+    is a fully connected layer with activation function.
 
     Parameters
     ----------

deepmd/dpmodel/fitting/polarizability_fitting.py

@@ -44,6 +44,30 @@
 class PolarFitting(GeneralFitting):
     r"""Fitting rotationally equivariant polarizability of the system.
 
+    The polarizability tensor :math:`\boldsymbol{\alpha} \in \mathbb{R}^{3 \times 3}` is
+    computed from the fitting network output and the rotation matrix:
+
+    **Diagonal fitting** (when `fit_diag=True`):
+
+    .. math::
+        \boldsymbol{\alpha}^i = \mathbf{R}^{i,T} \cdot \mathrm{diag}(\mathbf{p}^i) \cdot \mathbf{R}^i,
+
+    where :math:`\mathbf{p}^i \in \mathbb{R}^{m_1}` is the diagonal elements predicted by
+    the fitting network.
+
+    **Full matrix fitting** (when `fit_diag=False`):
+
+    .. math::
+        \boldsymbol{\alpha}^i = \mathbf{R}^{i,T} \cdot \mathbf{P}^i \cdot \mathbf{R}^i,
+
+    where :math:`\mathbf{P}^i \in \mathbb{R}^{m_1 \times m_1}` is the full matrix predicted
+    by the fitting network.
+
+    The fitting network is:
+
+    .. math::
+        \mathbf{p}^i \text{ (or } \mathbf{P}^i) = \mathcal{L}^{(n)} \circ \cdots \circ \mathcal{L}^{(0)}(\mathcal{D}^i).
+
     Parameters
     ----------
     ntypes