Avoid a graph break in ModuleDict and ParameterDict

[akihironitta]

We have a graph break at hasattr call in ModuleDict.to_internal_key.

Repro:

import torch
from torch_geometric.nn.module_dict import ModuleDict

edge_type = ("a", "to", "b")

class SomeModel(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.module_dict = ModuleDict({
            edge_type: torch.nn.Linear(1, 1),
        })

    def forward(self, x):
        # need to convert tuple to string in advance to avoid a graph break
        # due to https://github.com/pytorch/pytorch/issues/111551
        key = ModuleDict.to_internal_key(edge_type)
        x = self.module_dict[key](x)
        return x

from torch._dynamo.utils import CompileProfiler
model = SomeModel()
with CompileProfiler() as prof:
    model = torch.compile(model)
    model(torch.randn(1, 1))
    print(prof.report())

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (b1f8535) 88.80% compared to head (171e494) 88.41%.

❗ Current head 171e494 differs from pull request most recent head c2f1378. Consider uploading reports for the commit c2f1378 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #8363      +/-   ##
==========================================
- Coverage   88.80%   88.41%   -0.39%     
==========================================
  Files         475      475              
  Lines       28841    28838       -3     
==========================================
- Hits        25611    25497     -114     
- Misses       3230     3341     +111     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[akihironitta]

Converting this to draft as it's currently not a proper fix.

[akihironitta]

For the note, graph breaks produeced by hasattr here were likely due to pytorch/pytorch#111522.