Postpone dictionary encoding decision for starting null pages.

Discussion from the dev mailing list: [Fallback Encoding for Very Sparse or Sorted Datasets-Apache Mail Archives](https://lists.apache.org/thread/jkkxhknsxw8cdv8fxf1bstqdn5fzw2pl)

If beginning values are all nulls, the dictionary encoding is disabled when first data page is created. This loses the advantage of dictionary encoding if following values are good fit.

**Reporter**: [Gang Wu](https://issues.apache.org/jira/secure/ViewProfile.jspa?name=wgtmac) / @wgtmac
**Assignee**: [Gang Wu](https://issues.apache.org/jira/secure/ViewProfile.jspa?name=wgtmac) / @wgtmac

_{**Note**: *This issue was originally created as [PARQUET-2253](https://issues.apache.org/jira/browse/PARQUET-2253). Please see the [migration documentation](https://issues.apache.org/jira/browse/PARQUET-2502) for further details.*}