Fix: Regression in XML train plugin parsing

[AmberIsFrozen]

Introduced in d1ea76d

Since text node is also a type of child node, HasChildNodes would always be true even with the old way of specifying plugin (<plugin>path_to_plugin.dll</plugin>), which cause existing train XML to fail to load.

Changes it to look for child XML Elements instead, if not then fallback to the old way of loading plugin using InnerText.

[leezer3]

Thanks.

As a note, I'm going to junk all the XML parsers in favor of block format soon:
#1214

Doing it that way has initial pain, but makes the thing so much more maintainable; you can only make the same mistake once....