AtomicConv applies dropout during eval, so predict() is nondeterministic

Bug

AtomicConv.forward calls F.dropout(x, dropout) without passing training. F.dropout defaults to training=True, so it stays active after model.eval(). dropouts defaults to 0.5, and AtomConvModel.predict() runs through this path, so repeated predictions on the same input differ.

deepchem/models/torch_models/layers.py:

if dropout > 0:
    x = F.dropout(x, dropout)

Other modules guard this. fcnet.py uses if dropout > 0.0 and self.training, and the uncertainty path uses an explicit dropout_switch. AtomicConv has neither.

To Reproduce

1. Build an AtomConvModel with a nonzero dropouts value (0.5 is the default).
2. Fit it, then call predict on the same dataset twice.
3. The two prediction arrays differ.

Isolating the mechanism:

import torch, torch.nn as nn, torch.nn.functional as F
m = nn.Linear(8, 8).eval()
t = torch.ones(1, 8)
print(torch.equal(F.dropout(m(t), 0.5), F.dropout(m(t), 0.5)))
# False
print(torch.equal(F.dropout(m(t), 0.5, training=m.training),
                  F.dropout(m(t), 0.5, training=m.training)))
# True

Expected behavior

Dropout is disabled under eval(), so predict() returns the same values for the same input.

Environment

• OS: macOS 26.5.1
• Python version: 3.12.11
• DeepChem version: main @ 43e6f2e
• PyTorch version: 2.8.0
• Any other relevant information: same defect class as #5086 / PR #5087, which fixed ScScore.forward

Additional context

Both this and the layer chaining question in #5091 sit in the same loop. Filing separately since they are independent.