Arm backend: Stabilize VGF bilinear fp16 test

This patch fixes a flaky VGF no-quant Interpolate test for
upsample_bilinear2d fp16 inputs.

The neighboring VGF Upsample test already uses looser fp16 comparison
tolerances, while the Interpolate variant still used the default 1e-3
atol/rtol. The observed fp16 output difference can be slightly above
that default threshold.

Match the Upsample behavior by using 1e-2 atol/rtol for fp16 inputs
and keeping 1e-3 atol/rtol for other dtypes.

The change was verified with collection of the targeted VGF Interpolate
pytest cases.

Signed-off-by: Baris Demir <baris.demir@arm.com>
Change-Id: I83e62f2241d418100d3fa9df5883b2935bd751a1

Author: bdemirb

backends/arm/test/ops/test_upsample_bilinear2d.py

@@ -551,12 +551,21 @@ def test_upsample_bilinear2d_vec_vgf_no_quant_Upsample(test_data: torch.Tensor):
 @common.SkipIfNoModelConverter
 def test_upsample_bilinear2d_vec_vgf_no_quant_Interpolate(test_data: torch.Tensor):
     data, size, scale_factor, compare = test_data()
+    match data.dtype:
+        case torch.float16:
+            atol = 1e-2
+            rtol = 1e-2
+        case _:
+            atol = 1e-3
+            rtol = 1e-3
     pipeline = VgfPipeline[input_t1](
         Interpolate(size, scale_factor),
         (data,),
         aten_op,
         exir_op,
         quantize=False,
+        atol=atol,
+        rtol=rtol,
     )
     if not compare:
         pipeline.pop_stage(-1)