Delay RNN graph break until module call (#186614)

Dynamo intentionally graph breaks on nn.RNN, nn.GRU, and nn.LSTM when
allow_rnn is disabled. The old implementation raised the unsupported error
while wrapping the recurrent module attribute. For LOAD_ATTR/LOAD_METHOD
bytecode this goes through the generic step graph-break path, which can
compile the prefix but cannot build a continuation for the call site, so the
rest of the frame runs eagerly and post-RNN code is never compiled.

Return a DelayGraphBreakVariable for unsupported recurrent modules instead.
This keeps the attribute load traceable enough to reach the CALL instruction,
where Dynamo's regular call graph-break handling can run the recurrent module
eagerly and resume tracing the continuation. This matches the existing delayed
break pattern used for disabled optimized modules and avoids changing the
experimental allow_rnn=True behavior.

Add regression coverage for RNN, GRU, and LSTM modules that verifies Dynamo
compiles both the pre-recurrent and post-recurrent regions.

Fixes #141184
Generated by my agent

Test Plan:
- python test/dynamo/test_modules.py -k test_rnn_graph_break_resumes_after_call
- python test/export/test_export.py -k test_export_dynamo_config
- python test/dynamo/test_modules.py -k unsupportedmodule
- lintrunner -a

Pull Request resolved: https://github.com/pytorch/pytorch/pull/186614
Approved by: https://github.com/mlazos