Skip installing large aliased constant-folded weights

[micwill755]

Summary

• Skip installing large constant-folded tensors that only alias existing module weights, avoiding expensive cpu().contiguous() copies under offload_module_to_cpu.

PR 2 of 3 in a compile-startup stack; based on #4530.

Test plan

• Flux NVFP4 cold compile still succeeds with offload_module_to_cpu=True
• Large weight permutes remain in-graph (not installed as frozen params)
• VLA still folds materialized embedding constants and compiles (int32 indices)