LinkBuilder ignores non-temporal graph edge scales

## Summary

`LinkBuilder.forward` only assigns the temporal-edge term to `adj_filtered`. The user, anchor-user, and fallback/malicious terms are written as separate expression statements, so their learnable scale parameters do not affect the returned adjacency.

This occurs in both `model/ours/ACMIL_v2.py` and `model/ours/patchMIL_v20.py` at commit `a0b47382f08b108a17a39b343c4e58dbf2852db3`.

## Minimal reproduction

```python
import torch
from model.ours.ACMIL_v2 import LinkBuilder

layer = LinkBuilder(hidden_dim=2, adj_scale_init=0.5)
patches = torch.tensor([[[1., 0.], [2., 0.], [3., 0.], [4., 0.]]])
valid = torch.ones(1, 4, dtype=torch.bool)
graph = torch.zeros(1, 4, 4, 3, dtype=torch.bool)
graph[0, 0, 1, 0] = True # temporal
graph[0, 0, 2, 1] = True # same user
graph[0, 0, 3, 2] = True # anchor-user

layer(patches, valid, graph).square().sum().backward()
for name, parameter in layer.named_parameters():
print(name, parameter.grad)
```

Actual result: `scale_factor_time_adj` has a gradient, while `scale_factor_user_adj`, `scale_factor_anchor_user_adj`, and `scale_factor_malicious_adj` all have `None` gradients. The same result occurs with `model.ours.patchMIL_v20.LinkBuilder`.

## Expected behavior

All four relation-specific terms should participate in the adjacency calculation, so each scale can influence the result and receive gradients when its edge category is present.

## Impact

Graph-aware attention currently ignores same-user, anchor-user, and fallback relationships despite defining trainable scales for them. Training can therefore only learn the temporal-edge scale.

## Suggested fix

Wrap the four weighted terms in a single parenthesized expression in both implementations and add a regression test that exercises every edge category.

Start with LinkBuilder.forward in model/ours/ACMIL_v2.py and model/ours/patchMIL_v20.py at commit a0b47382f08b108a17a39b343c4e58dbf2852db3. Run the minimal PyTorch reproduction, then add a regression test covering temporal, same-user, anchor-user, and fallback/malicious edges. Done means all four relation-specific scale parameters affect the adjacency and receive gradients when their edge category is present.

Written by the indexing model from the issue text.