optimizer use `default_main_program()` as key for learning rate map, `Program`
seem to rely on object hash for equality check, and cloned program is not equal
to the original

parameters are bound to model not program, critical for transfer learning to work
parameters should be reusable between static/dynamic graph
optimizer state is not reusable

1. weights are bound to respective models
2. models can coexists in same scope (actual parameter names are uniqued)
3. this make model composition (e.g., transfer learning) harder