Update default LayerwiseScheduler behavior (#3291)

### Changes

`add_additional_outputs = True` is set in `LayerwiseScheduler` by
default. It's mean that additional output nodes will be added that were
visited to remove inputs to target nodes.
### Reason for changes

Memory and speed optimization. 

More details:
Calibration dataset: wikitext-2-raw-v1
```
compressed_model = nncf.compress_weights(
      model,
      dataset=quantization_dataset,
      mode=nncf.CompressWeightsMode.INT4_SYM,
      gptq=True,
      subset_size=128,
      ratio=1,
  )
```

1. sequence length > 128
- Reduce memory footprint in 2.71 times.
- 1.16x compression time speed up. 

Model | Accuracy (lambada_openai) Develop | Accuracy (lambada_openai)
Branch | Compression time Develop (sec.) | Compression time Branch
(sec.) | Compression time speed up | Peak Memory Develop (MiB) | Peak
Memory Branch (MiB) | Memory reduction
-- | -- | -- | -- | -- | -- | -- | -- | --
facebook/opt-125m | аcc: 0.3369 perplexity: 32.2038 | аcc: 0.3369
perplexity: 32.2038 | 219 | 204 | 1,07x | 5575 | 2341 | 2,38x
TinyLlama/TinyLlama-1.1B-Chat-v1.0 | аcc: 0.5956 perplexity: 6.7443 |
аcc: 0.5956 perplexity: 6.7443 | 1889 | 1620 | 1,17x | 25980 | 8413 |
3,09x
microsoft/Phi-3.5-mini-instruct | аcc: 0.6328 perplexity: 5.6143 | аcc:
0.6328 perplexity: 5.6143 | 7994 | 6334 | 1,26x | 61273 | 22968 | 2,67x

2. sequence length = 32
- Reduce memory footprint in 1.42 times.
- 1.07x compression time speed up. 

Model | Compression time Develop (sec.) | Compression time Branch (sec.)
| Compression time speed up | Peak Memory Develop (MiB) | Peak Memory
Branch (MiB) | Memory reduction
-- | -- | -- | -- | -- | -- | --
facebook/opt-125m | 192 | 183 | 1,05 | 2205 | 1695 | 1,30
TinyLlama/TinyLlama-1.1B-Chat-v1.0 | 1596 | 1482 | 1,08 | 9247 | 6182 |
1,50
microsoft/Phi-3.5-mini-instruct | 6724 | 6196 | 1,09 | 25924 | 17765 |
1,46



Memory profiler for TinyLlama/TinyLlama-1.1B-Chat-v1.0:
Develop:

![image](https://github.com/user-attachments/assets/2a819b19-89c3-434e-8751-74911a41348e)
Branch:

![image](https://github.com/user-attachments/assets/501652fb-6865-4484-9cc6-149d0899d9a8)

Memory profiler for microsoft/Phi-3.5-mini-instruct:
Develop:

![image](https://github.com/user-attachments/assets/ba64e5ce-36eb-4b57-a35b-9199bfcc4e44)
Branch:

![image](https://github.com/user-attachments/assets/4f742016-0ceb-42f7-bd04-53f7d112b789)

### Related tickets

ref: 153732

### Tests

NNCF/job/manual/job/post_training_weight_compression/329/

Author: alexsu52

nncf/quantization/algorithms/layerwise/scheduler.py

@@ -82,9 +82,10 @@ class LayerwiseScheduler:
     7. Repeat the process until all target nodes have been processed.
     """
 
-    def __init__(self, add_additional_outputs: bool = False):
+    def __init__(self, add_additional_outputs: bool = True):
         """
-        :param strategy: The strategy to use for scheduling.
+        :param add_additional_outputs: If True (default), includes additional output nodes that were visited
+            to remove inputs to target nodes.
         """
         self.add_additional_outputs = add_additional_outputs