Fix xsd:import without namespace incorrectly inheriting parent targetNamespace

[pauloxnet]

The block at lines 255-259 of src/zeep/xsd/visitor.py (elif not schema_tns and not namespace: namespace = self.document._target_namespace) forces an xsd:import schemaLocation="..." without a namespace attribute to inherit the importing schema's targetNamespace. That is xsd:include semantics, not xsd:import semantics. The practical effect: elements declared in the imported no-namespace schema get re-registered under the parent namespace, and any unqualified QName referencing them (typically wsdl:part element="Header" in a soap:header binding) fails to resolve, surfacing as IndexError: No definition '{tns}Header' in 'messages' found. WCF/.NET-generated WSDLs hit this routinely.

The block was introduced in 713c779 ("Fix an issue with resolving qualified qname's used in references", #879, Dec 2018). One week later #888 / b7ad93e partially reverted the chameleon support and added the _has_empty_import flag, but only for the <import/> case (no namespace, no location). At the time @mvantellingen noted "I'm wondering if that unittest is actually ok" and @apollo13 replied "Not sure if the test is wrong or the code needs more work". That uncertainty was never resolved. Issues #600 (Jan 2018) and #662 (Jun 2018) report the same symptom against WCF WSDLs and have stayed open or closed-without-fix since.

Fix

Two changes in visit_import:

1. The inheritance block is removed.
2. _has_empty_import is activated whenever an import omits namespace, per W3C XSD 4.1.2. The activation is placed at the three exit points so that an import naming a namespace incorrectly (which the existing XMLParseError rejects a few lines later) cannot poison the parent's QName resolution: on a cache hit, only when the cached document has no _target_namespace; on the canonical empty <import/>, in the missing-schemaLocation branch; and post-load, only when the imported schema has no targetNamespace either.

tests/test_wsdl.py::test_import_schema_without_namespace_and_header reproduces the original failure (namespaced schema importing a no-namespace schema whose element is used as a SOAP header part) and asserts both that the element ends up under the null namespace and that no {tns}Header ghost registration exists.

Breaking change

Strict W3C QName resolution has a real consequence. In a schema with elementFormDefault="qualified" that does <xsd:import schemaLocation="..."/> (no namespace) of a no-namespace schema, an unqualified ref to an element defined in the same importing schema now resolves in the null namespace and fails. The spec-compliant fix on the consumer side is the qualified form:

<xsd:element ref="tns:local"/>

This case is captured in tests/test_xsd_schemas.py::test_no_target_namespace_unqualified_ref_to_own_element, marked xfail(strict=True) so any future change is a deliberate one. No existing test in the repository hits this pattern.