Stack too deep when subnetworks are involved

[albertz]

With subnetworks, the stack can become quite deep and convoluted (independent of not using flat net construction, #992), like:

_create_layer, network.py:1048
add_layer, network.py:1118
add_layer, network.py:1116
add_layer, network.py:1116
add_layer, network.py:1116
construct_layer, network.py:959
get_layer, network.py:839
wrapped_get_layer, network.py:3011
wrapped_get_layer, network.py:3030
wrapped_get_layer, network.py:3011
wrapped_get_layer, network.py:3030
wrapped_get_layer, network.py:3011
wrapped_get_layer, network.py:3030
<listcomp>, base.py:589
transform_config_dict, base.py:588
construct_layer, network.py:952
get_layer, network.py:839
wrapped_get_layer, network.py:3011
wrapped_get_layer, network.py:3030
wrapped_get_layer, network.py:3011
wrapped_get_layer, network.py:3030
wrapped_get_layer, network.py:3011
wrapped_get_layer, network.py:3030
...
_construct_template, rec.py:1569
__init__, rec.py:1180
transform_config_dict, rec.py:406
construct_layer, network.py:952
get_layer, network.py:839
construct_layer, network.py:919
get_layer, network.py:839
<listcomp>, base.py:589
transform_config_dict, base.py:588
transform_config_dict, basic.py:363
construct_layer, network.py:952
construct_from_dict, network.py:633
create_network, engine.py:1342
_init_network, engine.py:1301
init_network_from_config, engine.py:1121
init_train_from_config, engine.py:1056

See here for a better stack trace: https://gist.github.com/albertz/a1a20710e4d08292e35ba6910ca132e8 (via #1001)

Esp the logic using wrapped_get_layer contributes to this.

This should be cleaned up.