Improve annotation of fromstring

`parser` seems like it can be an `HTMLParser. There's code in the wild
that does this without obviously being wrong. As far as I can tell from
the lxml Cython source, `parser` is a `_Baseparser`. I wasn't sure if
the `iterparse` class ought to be allowed here too. I erred on the side
of caution.

It's possible for this function to return None, if the parser has
`recover=True`. I couldn't find a good way to express this without
affecting every other call to `fromstring`. (Perhaps one could make the
Parser generic over some kind of `Recoverable` indicator type... but
that seemed like overkill)

Resolves lxml#63.

lxml-stubs/etree.pyi

@@ -496,8 +496,11 @@ def parse(
     source: _FileSource, parser: XMLParser = ..., base_url: _AnyStr = ...
 ) -> _ElementTree: ...
 def fromstring(
-    text: _AnyStr, parser: XMLParser = ..., *, base_url: _AnyStr = ...
-) -> _Element: ...
+    text: _AnyStr,
+    parser: Union[XMLParser, HTMLParser] = ...,
+    *,
+    base_url: _AnyStr = ...,
+) -> Optional[_Element]: ...
 @overload
 def tostring(
     element_or_tree: _ElementOrTree,

test-data/test-etree.yml

@@ -3,7 +3,15 @@
     main: |
         from lxml import etree
         document = etree.fromstring("<doc></doc>")
-        reveal_type(document)  # N: Revealed type is "lxml.etree._Element"
+        reveal_type(document)  # N: Revealed type is "Union[lxml.etree._Element, None]"
+
+-   case: etree_from_empty_string_with_parser_recovery_returns_none
+    disable_cache: true
+    main: |
+        from lxml import etree
+        parser = etree.HTMLParser(recover=True)
+        document = etree.fromstring("", parser)
+        reveal_type(document)  # N: Revealed type is "Union[lxml.etree._Element, None]"
 
 -   case: etree_element_find
     disable_cache: true