avoid duplicate Content-Length header in DefaultClient

[alhudz]

Repro: send a body-less POST/PUT whose template already carries a Content-Length header (e.g. @Headers("Content-Length: 0")), with sun.net.http.allowRestrictedHeaders=true so the header reaches the wire.
Cause: DefaultClient.convertAndSend writes the request Content-Length in the header loop and then unconditionally adds a second Content-Length: 0 for every body-less method that allows a body, so the field goes out twice. RFC 7230 §3.3.2 forbids generating multiple Content-Length fields; the duplicate is ambiguous framing and servers reject it (Tomcat returns 400, #2862).
Fix: only add the fallback Content-Length: 0 when the request does not already declare a Content-Length.

The regression test is gated on sun.net.http.allowRestrictedHeaders like the existing Content-Length tests, since HttpURLConnection otherwise drops the restricted header, and asserts the field is sent once.

[alhudz]

any update?

[velo]

Thanks for the fix and the wire-level regression test! We ended up merging #3448, which removes the manual Content-Length write from the header loop entirely and lets the JDK own the header — that also covers the bodyless-request duplicate this PR guards against (the "0" fallback now fires exactly once). Holding off on merging this one to avoid stacking two untested-in-combination changes to the same code path; if you think the containsKey guard is still needed on top of #3448, happy to revisit.

[alhudz]

Checked this against current master with #3448 in, by running my gated regression test on top of master without the containsKey guard:

repro: body-less POST with @Headers("Content-Length: 0") and sun.net.http.allowRestrictedHeaders=true
before #3448: recorded Content-Length = ["0", "0"]
on master now: recorded Content-Length = ["0"], test passes with no DefaultClient change

So agreed, the guard isn't needed any more. The header loop no longer writes Content-Length, which leaves the body-less 0 fallback as the only writer. Fine to close this one. If wire-level coverage for the #2862 case is worth keeping (the test in #3448 is mock-level and covers the body path), I can strip this down to just the gated test.