[OpenVINO] Add HW patterns (#2822)

### Changes

Add HW patterns covering mask_rcnn_resnet50_atrous_coco model.
From an execution graph perspective, the subgraph of a model with a new
quantization scheme executes up to 3x faster.
Before | After

![image](https://github.com/user-attachments/assets/c94df9db-9b05-4ffa-9a30-4c93cc5127f1)


### Reason for changes
Improve performance. 

Local measurements:
Before: 1.8666 FPS | After: 2.0266 FPS 
Speed up ~ 8-9%
### Related tickets

144218

### Tests

Manually tested on the model.

nncf/common/graph/patterns/patterns.py

@@ -350,6 +350,9 @@ class HWFusedPatternNames(Enum):
     LINEAR_ARITHMETIC_ACTIVATIONS_ARITHMETIC = PatternDesc("linear_arithmetic_activations_arithmetic")
     LINEAR_BATCH_NORM = PatternDesc("linear_batch_norm")
     LINEAR_BATCH_NORM_ACTIVATIONS = PatternDesc("linear_batch_norm_activations")
+    # MaskRCNN_Resnet_Atrous
+    LINEAR_BATCH_TO_SPACE_SCALE_SHIFT_ACTIVATIONS = PatternDesc("linear_batch_to_space_scale_shift_activations")
+    LINEAR_BATCH_TO_SPACE_ARITHMETIC_ACTIVATIONS = PatternDesc("linear_batch_to_space_arithmetic_activations")
     LINEAR_BATCH_NORM_SCALE_SHIFT_ACTIVATIONS = PatternDesc("linear_batch_norm_scale_shift_activations")
     LINEAR_SCALE_SHIFT_ACTIVATIONS = PatternDesc("linear_scale_shift_activations")
     LINEAR_CONST_MULTIPLY = PatternDesc("linear_const_multiply")

nncf/openvino/hardware/fused_patterns.py

@@ -666,6 +666,32 @@ def create_linear_biased_activation_scale_shift() -> GraphPattern:
     return linear_biased
 
 
+@OPENVINO_HW_FUSED_PATTERNS.register(HWFusedPatternNames.LINEAR_BATCH_TO_SPACE_SCALE_SHIFT_ACTIVATIONS)
+def create_linear_batch_to_space_scale_shift_activations() -> GraphPattern:
+    linear = linear_operations()
+    batch_to_space = batch_to_space_operation()
+    scale_shift = create_scale_shift()
+    activations = atomic_activations_operations()
+
+    linear.join_patterns(batch_to_space)
+    linear.join_patterns(scale_shift)
+    linear.join_patterns(activations)
+    return linear
+
+
+@OPENVINO_HW_FUSED_PATTERNS.register(HWFusedPatternNames.LINEAR_BATCH_TO_SPACE_ARITHMETIC_ACTIVATIONS)
+def create_linear_batch_to_space_arithmetic_activations() -> GraphPattern:
+    linear = linear_operations()
+    batch_to_space = batch_to_space_operation()
+    arithmetic = arithmetic_operations()
+    activations = atomic_activations_operations()
+
+    linear.join_patterns(batch_to_space)
+    linear.join_patterns(arithmetic)
+    linear.join_patterns(activations)
+    return linear
+
+
 @OPENVINO_HW_FUSED_PATTERNS.register(HWFusedPatternNames.LINEAR_ELEMENTWISE)
 def create_linear_elementwise() -> GraphPattern:
     linear = linear_operations()
@@ -748,6 +774,14 @@ def unsqueeze_operation() -> GraphPattern:
     return pattern
 
 
+def batch_to_space_operation() -> GraphPattern:
+    pattern = GraphPattern()
+    pattern.add_node(
+        **{GraphPattern.LABEL_ATTR: "BATCH_TO_SPACE", GraphPattern.METATYPE_ATTR: om.OVBatchToSpaceMetatype}
+    )
+    return pattern
+
+
 def create_input_convert_transpose() -> GraphPattern:
     pattern = GraphPattern()
     model_input = pattern.add_node(

tests/onnx/test_pattern_manager.py

@@ -46,6 +46,8 @@
     HWFusedPatternNames.MVN_SCALE_SHIFT_ACTIVATIONS: "Not relevant for ONNX.",
     HWFusedPatternNames.LINEAR_ACTIVATIONS_UNSQUEEZE_BN_SQUEEZE: "Not relevant for ONNX.",
     HWFusedPatternNames.ARITHMETIC_ACTIVATIONS_ARITHMETIC: "Not relevant for ONNX.",
+    HWFusedPatternNames.LINEAR_BATCH_TO_SPACE_ARITHMETIC_ACTIVATIONS: "Not relevant for ONNX.",
+    HWFusedPatternNames.LINEAR_BATCH_TO_SPACE_SCALE_SHIFT_ACTIVATIONS: "Not relevant for ONNX.",
 }
 
 IGNORING_IGNORED_PATTERN_REASONS = {

tests/torch/test_pattern_manager.py

@@ -67,6 +67,8 @@
     HWFusedPatternNames.MVN: "Not relevant for Torch.",
     HWFusedPatternNames.GELU: "Not relevant for Torch.",
     HWFusedPatternNames.ARITHMETIC_ACTIVATIONS_ARITHMETIC: "Not relevant for Torch.",
+    HWFusedPatternNames.LINEAR_BATCH_TO_SPACE_ARITHMETIC_ACTIVATIONS: "Not relevant for Torch.",
+    HWFusedPatternNames.LINEAR_BATCH_TO_SPACE_SCALE_SHIFT_ACTIVATIONS: "Not relevant for Torch.",
 }
 
 IGNORING_IGNORED_PATTERN_REASONS = {