VariableLayer, change default add_batch_axis to False

[albertz]

Fix #763.

As this should not have any influence on behavior, this does not need a new behavior version.

[albertz]

The assumption that it should not have any influence (or rather that all other layers are bug free and feature complete) was invalid. (@Zettelkasten )

The test test_ScatterNdLayer_RangeLayer_RangeInAxisLayer fails.
Network:

net_dict = {
  "t": {
    "class": "range_in_axis", "axis": "t", "from": "data",
    "out_shape": {time_dim, ImplicitDynSizeDim(BatchDim)}},  # (T,)
  "range": {"class": "range", "limit": n_ts, "out_spatial_dim": ts_dim},  # (Ts,)
  "add_t": {
    "class": "combine", "kind": "add", "from": ["t", "range"],
    "out_shape": {time_dim, ts_dim, ImplicitDynSizeDim(BatchDim)}},  # (T,Ts)
  "t_rel_var": {"class": "variable", "shape": (ts_dim, n_out), "init": "glorot_uniform"},  # (B,Ts,D)
  "output": {"class": "scatter_nd", "from": "t_rel_var", "position": "add_t", "position_axis": -1,
             "output_dim_via_time_from": "data", "filter_invalid_indices": True}
}

Log:

layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'t_rel_var' output: Data{'t_rel_var_output', ['ts'(6),F|F't_rel_var:static:1'(11)]}
layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'data' output: Data{'data', [B,T|'time'[B],F|F'in-feature'(7)]}
2021-11-24 15:15:03.524447: W tensorflow/python/util/util.cc:348] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them.
layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'t' output: Data{'t_output', [T|'time'[B]], dtype='int32'}
layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'range' output: Data{'range_output', [F|'ts'(6)], dtype='int32'}
layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'add_t' output: Data{'add_t_output', [T|'time'[B],F|'ts'(6)], dtype='int32'}
layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'output' output: Data{'output_output', [T|'time'[B],'time'[B?],F|F't_rel_var:static:1'(11)]}
EXCEPTION
Traceback (most recent call last):
  File "tests/test_TFNetworkLayer.py", line 5634, in <module>
    line: globals()[arg]()  # assume function and execute
    locals:
      globals = <builtin> <built-in function globals>
      arg = <local> 'test_ScatterNdLayer_RangeLayer_RangeInAxisLayer', len = 47
  File "tests/test_TFNetworkLayer.py", line 2509, in test_ScatterNdLayer_RangeLayer_RangeInAxisLayer
    line: assert out_layer.output.feature_dim_axis_or_unspecified is NotSpecified and out_layer.output.feature_dim_axis == 3
    locals:
      out_layer = <local> <ScatterNdLayer 'output' out_type=Data{[T|'time'[B],'time'[B?],F|F't_rel_var:static:1'(11)]}>
      out_layer.output = <local> Data{'output_output', [T|'time'[B],'time'[B?],F|F't_rel_var:static:1'(11)]}
      out_layer.output.feature_dim_axis_or_unspecified = <local> <class 'returnn.util.basic.NotSpecified'>
      NotSpecified = <global> <class 'returnn.util.basic.NotSpecified'>
      out_layer.output.feature_dim_axis = <local> 2
AssertionError

I assume some problem in ScatterNdLayer. I will look into that.

[albertz]

Ok, maybe just the test is a bit strange.
For reference, the layer log in current master:

layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'t_rel_var' output: Data{'t_rel_var_output', [B,'t_rel_var:static:0'(6),F|F't_rel_var:static:1'(11)]}
layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'data' output: Data{'data', [B,T|'T'[B],F|F'F'(7)]}
2021-11-24 15:56:34.763191: W tensorflow/python/util/util.cc:348] Sets are not currently considered sequences, but this may change in the future, so consider avoiding using them.
layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'t' output: Data{'t_output', [T|'T'[B]], dtype='int32'}
layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'range' output: Data{'range_output', [F|'ts'(6)], dtype='int32'}
layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'add_t' output: Data{'add_t_output', [T|'T'[B],F|'ts'(6)], dtype='int32'}
layer <network via test_ScatterNdLayer_RangeLayer_RangeInAxisLayer>/'output' output: Data{'output_output', [T|'T'[B],'T'[B],B,F|F't_rel_var:static:1'(11)]}

It looks correct that we have the T axis twice in the output.

This is a case where enforced unique dim tags (#632) would fail.