fix: reject a message whose Transfer-Encoding value cannot be parsed

[pjfanning]

cherry pick 44b01c3 #1267

Motivation:
When a modelled header value fails to parse, ModeledHeaderValueParser calls onIllegalHeader - which by default only logs - and degrades the header to a RawHeader. For Transfer-Encoding that is unsafe: the degraded header never reaches the case h: Transfer-Encoding arm in parseHeaderLines, so isChunked stays false and, with a Content-Length also present, the message is framed by Content-Length. An upstream that does parse the value - stripping quotes, tolerating trailing junk - frames the same message as chunked. The two disagree about where the message ends, which is a request smuggling primitive.

Pekko HTTP already rejects the clear cases: chunked together with a Content-Length, an unsupported coding, and multiple entries. Only the unparseable value was silently tolerated.

Modification:
Fail the message when a Transfer-Encoding header arrives as a RawHeader. That can only happen when the modelled parse failed: transfer-encoding is in alwaysParsedHeaders, so it is modelled even when modeled-header-parsing is off, and a well-formed value always reaches the modelled arm.

Result:
A message whose Transfer-Encoding cannot be understood is rejected rather than framed by a different rule than the sender used.

Tests:

• sbt "http-core/testOnly org.apache.pekko.http.impl.engine.parsing.*" - pass (246 tests); a new test sends Transfer-Encoding: "chunked" alongside a Content-Length and expects a 400. Verified it fails without the change, where the message is accepted and framed by Content-Length.
• sbt http-core/mimaReportBinaryIssues - pass

References:
None - rejects a message whose Transfer-Encoding value cannot be parsed