hoist precomputed exprs from indices (#95690)
This generates compilable code for maskrcnn graph 13, with ceilings hoisted to be computed on the host. But it now fails with
```
File "/scratch/ngimel/work/pytorch/torch/_dynamo/symbolic_convert.py", line 379, in wrapper
self.output.compile_subgraph(self, reason=reason)
File "/scratch/ngimel/work/pytorch/torch/_dynamo/output_graph.py", line 562, in compile_subgraph
pass1.foreach(stack_values)
File "/scratch/ngimel/work/pytorch/torch/_dynamo/codegen.py", line 166, in foreach
self(i)
File "/scratch/ngimel/work/pytorch/torch/_dynamo/codegen.py", line 148, in __call__
output.extend(value.reconstruct(self))
File "/scratch/ngimel/work/pytorch/torch/_dynamo/variables/dicts.py", line 40, in reconstruct
codegen.create_load_python_module(collections),
TypeError: create_load_python_module() missing 1 required positional argument: 'push_null'
from user code:
File "/scratch/ngimel/work/env/lib/python3.9/site-packages/torchvision-0.15.0a0+928b05c-py3.9-linux-x86_64.egg/torchvision/models/detection/backbone_utils.py", line 58, in forward
x = self.fpn(x)
```
looks like we never execute this `create_load_python_module()` path for other subgraphs.
Any advice on how to fix this @voznesenskym @jansel ?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/95690
Approved by: https://github.com/jansel
