random data is generated when output dtype=np.float16?

If module returns a 2D float16 Tensor with **1st dimension size > 5** and **2nd dimension size > 12** and **2nd dimension size is not multiple of 32**, it returns some random data from MLModel.predict. Do I need to set some other options to make output conversion correct?

example script to reproduce it
```python
import torch
import numpy as np
import coremltools as ct

class MyModule(torch.nn.Module):
def __init__(self):
super().__init__()

def forward(self, x: torch.Tensor):
return torch.ones(5, 13)

module = MyModule()
module.eval()

x = torch.ones((1), dtype=torch.float16)

traced_module = torch.jit.trace(module, [x])

mlmodel = ct.convert(
traced_module,
convert_to="mlprogram",
inputs=[ct.TensorType(name="x", shape=x.shape, dtype=np.float16)],
outputs=[ct.TensorType(name="output", dtype=np.float16)],
compute_units=ct.ComputeUnit.ALL,
compute_precision=ct.precision.FLOAT16,
minimum_deployment_target=ct.target.macOS13,
)

torch_output = traced_module(x)
print("torch output:", torch_output)

coreml_output = torch.from_numpy(
mlmodel.predict({'x': x})['output']
)
print(f"coreml output:", coreml_output)

diff = torch.abs(torch_output - coreml_output)
print("diff avg:", torch.mean(diff), ", diff max:",torch.max(diff))
```

example result from console. The result is different in each run.
```
Converting PyTorch Frontend ==> MIL Ops: 88%|███████████████████████▋ | 7/8 [00:00<00:00, 6273.53 ops/s]
Running MIL frontend_pytorch pipeline: 100%|█████████████████████████| 5/5 [00:00<00:00, 17549.39 passes/s]
Running MIL default pipeline: 100%|████████████████████████████████| 64/64 [00:00<00:00, 11410.16 passes/s]
Running MIL backend_mlprogram pipeline: 100%|██████████████████████| 11/11 [00:00<00:00, 67948.96 passes/s]
torch output: tensor([[1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.],
[1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1., 1.]])
coreml output: tensor([[1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00,
1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00,
1.0000e+00],
[1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00,
1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00, 1.0000e+00,
1.0000e+00],
[1.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00, 6.8528e-36, 1.4013e-45,
1.0693e-09, 1.4013e-45, 4.2039e-45, 0.0000e+00, 0.0000e+00, 0.0000e+00,
0.0000e+00],
[9.0943e-10, 1.4013e-45, 0.0000e+00, 0.0000e+00, 1.0711e-09, 1.4013e-45,
0.0000e+00, 0.0000e+00, 0.0000e+00, 1.4013e-45, 0.0000e+00, 0.0000e+00,
2.5644e-43],
[0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00,
0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00, 0.0000e+00,
0.0000e+00]])
diff avg: tensor(0.5846) , diff max: tensor(1.)
```
---
testing environment:
coremltools 7.0b1
torch 2.0.1
numpy 1.25.0
macOS Ventura 13.4 on Macbook M1 Air

Start by running the provided conversion script and comparing traced_module(x) with mlmodel.predict({'x': x})['output'] for the reported float16 shapes. Trace the conversion and prediction path responsible for the output buffer; done means the converted model consistently returns the expected all-ones 5×13 tensor without random values.

Written by the indexing model from the issue text.