Array input with integers results in "value type not convertible"

Repro steps:

```python
import coremltools
import numpy as np

spec = coremltools.proto.Model_pb2.Model()
spec.specificationVersion = 1
spec.identity.MergeFromString(b'')
input = spec.description.input.add()
input.type.multiArrayType.shape.append(3)
input.type.multiArrayType.dataType = coremltools.proto.FeatureTypes_pb2.ArrayFeatureType.INT32
input.name = "input"
output = spec.description.output.add()
output.type.multiArrayType.shape.append(3)
output.type.multiArrayType.dataType = coremltools.proto.FeatureTypes_pb2.ArrayFeatureType.INT32
output.name = "output"
model = coremltools.models.MLModel(spec)
model.predict({'input': [1,2,3]})
```

Expected: something like
```python
{'input': np.array([ 1., 2., 3.])}
```

Actual:
```python
/Users/zach/venv/lib/python2.7/site-packages/coremltools/models/model.pyc in predict(self, data, useCPUOnly, **kwargs)
318
319 if self.__proxy__:
--> 320 return self.__proxy__.predict(data,useCPUOnly)
321 else:
322 if _macos_version() < (10, 13):

RuntimeError: value type not convertible
```

Note that changing the input to `[1.0, 2, 3]` seems to fix the issue; so despite the multiArrayType being INT32, it only seems to allow float input (at least in some cases).

Start at coremltools.models.MLModel.predict, where the reported traceback reaches model.pyc, and reproduce the provided INT32 multi-array example. Trace why an integer list is rejected while a list containing a float succeeds, then add regression coverage showing that integer input is accepted for an INT32 array without the runtime error.

Written by the indexing model from the issue text.