Can't Convert Keras Model with Flexible Input Length and Dilated Convolution(s)

## ❓Question
I have a keras model that contains some dilated convolutions and whose input shape contains a variable component. When trying to convert the model using `coremltools.convert()`, I get the following error: `Not support dynamic block_shape and paddings for SpaceToBatchND!`. If I specify the variable component in the input shape, I get this error instead: `non-equal block shape is not yet supported for 4d input.`. I managed to create two very simple examples that showcase these two problems:

1. Variable Input Length
```python
import tensorflow as tf
import coremltools as ct

inputs = tf.keras.layers.Input(shape=(1, None, 1))
dilated_conv = tf.keras.layers.Conv1D(filters=20, kernel_size=5, dilation_rate=2)(inputs)
model = tf.keras.Model(inputs=inputs, outputs=[dilated_conv])

ct.convert(model) # Fails with error: "Not support dynamic block_shape and paddings for SpaceToBatchND!"
```

2. Fixed Input Length
```python
import tensorflow as tf
import coremltools as ct

inputs = tf.keras.layers.Input(shape=(1, 300, 1))
dilated_conv = tf.keras.layers.Conv1D(filters=20, kernel_size=5, dilation_rate=2)(inputs)
model = tf.keras.Model(inputs=inputs, outputs=[dilated_conv])

ct.convert(model) # Fails with error: "non-equal block shape is not yet supported for 4d input."
```

I've done some extensive debugging but I have no idea how to approach this. Initially, padding was set to 'same', which I thought may have been causing the issue; so I tried adding a padding layer before the convolution instead of specifying it when creating the layer. But as you can see from the examples above, it fails even if padding isn't specified at all.
I would appreciate any help on this issue! :)

## System Information
`coremltools.__version__ == '5.1.0'`
`tensorflow.__version__=='2.5.0'`

Start by running the two minimal Keras and coremltools conversion examples against the reported TensorFlow 2.5.0 and coremltools 5.1.0 versions, then trace the reported SpaceToBatchND errors in the converter. Done means both variable-length and fixed-length dilated-convolution models convert successfully, with regression coverage for the two cases.

Written by the indexing model from the issue text.