NXP backend: Fix bug with reused Linear bias. (#23112)

### Summary
Sometimes, the bias of multiple `Linear` nodes can be shared, so all
occurrences have the same quantization parameters. But as the bias's
scale is determined by the scale of the input and weight, an error is
introduced. This commit assigns each bias the correct scale during
conversion to Neutron IR. By doing so, 7 LSTM harness tests are enabled,
leaving only 3 failing ones (2 timeouts and 1 accuracy issue. Fixing
them is not a priority as proper LSTM support is planned for the
future).

### Test plan
`pytest -c /dev/nul backends/test /suite/operators/test_lstm.py -m
backend_nxp`


cc @robert-kalmar @JakeStevens @digantdesai @rascani

Author: MartinPavella

backends/nxp/backend/ir/converter/node_converters/ops_converters/addmm_converter.py

@@ -3,6 +3,7 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+import numpy as np
 import torch
 
 from executorch.backends.nxp.backend.edge_helper import (
@@ -15,6 +16,7 @@
     CustomDelegationOptions,
     NodeConverter,
 )
+from executorch.backends.nxp.backend.ir.tflite_generator import tflite_model
 from executorch.backends.nxp.backend.ir.tflite_generator.builtin_options import (
     fully_connected_options,
 )
@@ -138,4 +140,21 @@ def convert(self, node: Node):
         #  be present in the output model if the weights are static.)
         ops.add_pre(self.builder.create_transpose_operator_before(t_op, 1, [1, 0]))
 
+        # Neutron IR requires the bias scale to be the product of the input and weight scale. During quantization the
+        #  bias may have been shared across ops, so its annotated scale can be incorrect.
+        #  Recompute the bias scale now that the tensor is unique per op.
+        if (
+            bias.quantization is not None
+            and x.quantization is not None
+            and w.quantization is not None
+        ):
+            expected_scale = np.array(x.quantization.scale.vector) * np.array(
+                w.quantization.scale.vector
+            )
+            if not np.allclose(bias.quantization.scale.vector, expected_scale):
+                bias.quantization = tflite_model.Quantization(
+                    scale=tflite_model.Scale(list(expected_scale)),
+                    zero_point=tflite_model.ZeroPoint([0] * len(expected_scale)),
+                )
+
         self.builder.append_operators(ops.flatten())

backends/test/suite/flows/nxp.py

@@ -31,16 +31,9 @@
 # Tests known to fail on Neutron due to known bugs. Marked as xfail
 # (strict=True) so that an unexpected pass is also reported.
 _NEUTRON_XFAILS = [
-    "test_lstm_batch_first_false",
-    "test_lstm_return_hidden_states",
-    "test_lstm_batch_sizes",
-    "test_lstm_seq_lengths",
-    "test_lstm_bidirectional",
-    "test_lstm_with_dropout",
-    "test_lstm_dtype",
-    "test_lstm_with_initial_states",
-    "test_lstm_feature_sizes",
-    "test_lstm_num_layers",
+    "test_lstm_return_hidden_states",  # Accuracy error.
+    "test_lstm_seq_lengths",  # Timeout.
+    "test_lstm_num_layers",  # Timeout.
 ]