Add a new field attribute "pointer" to indicate that the generated
client code for the field must be a pointer. This allows clients to
differentiate between sending nil/leaving the value unset and sending an
empty map or slice.
This change also removes the `nullablestring` norman type introduced in
30f8d18 since schemas that need a pointer to a string can now use this
field attribute. There are no libraries currently using this feature so
it should be safe to remove.
Example usage:
```
Labels map[string]string `json:"labels" norman:"pointer"`
```
Resulting API schema:
```
"labels": {
"create": true,
"nullable": true,
"pointer": true,
"type": "map[string]",
"update": true
}
```
Generated client code:
```
Labels *map[string]string `json:"labels,omitempty" yaml:"labels,omitempty"`
```
Currently, `*string` fields in Go structs are converted to `{type:
string, nullable: true}` in the API schema, which is right, but then
converted down to just `string` when converted to generated client
structs. Without this patch, there is no way to express that `*string`s
should stay as `*string`s when run through the generator, because the
'nullable' flag is ignored. Compare this to `*bool`s, which become
`boolean` in the schema and then correctly converted back to *bool in
the generator. This patch adds a backwards-compatible way to opt in to
converting `*string`s in the original struct to *strings in the
generated struct.
Prior, empty array were converted to nil. In some cases the
difference between a nil and empty array are not arbitrary.
Now, empty arrays will be converted to empty interface arrays
and nil array will be converted to nil interface arrays.
