fix: don't fail a notification POST on response shape

[otherwhere666]

Problem

A POST that carries only notifications/responses/errors leaves no request outstanding, so nothing in the response is load-bearing. The spec says the server SHOULD reply 202 Accepted with no body — but servers in the wild routinely reply 200 instead, with or without a body, and with or without a Content-Type.

post_message currently accepts such a POST only on 202/204, on a 2xx carrying a literal Content-Length: 0, or on a 2xx typed application/json / text/event-stream. Anything else returns UnexpectedContentType, which fails the send. When the message is notifications/initialized, that takes the whole handshake down against a server whose only fault is answering 200 instead of 202.

Two shapes hit this in practice:

1. 200 with an unexpected or absent Content-Type. The TypeScript SDK gates its content-type validation behind hasRequests, so a notification-only POST skips the check entirely and any 2xx passes. We are stricter than the other client SDKs a server is likely to have been tested against — a server can pass every other client and fail only here.
2. An empty body sent with chunked transfer encoding. Content-Length is absent, so content_length == Some(0) cannot fire even though the body is in fact empty, and the request falls through to the fatal arm.

We hit (1) against a stateless server that returned 200 for notifications/initialized. The same endpoint completed the handshake against five other MCP clients; ours was the only one that rejected it.

Change

Never fail a POST that awaits no response on response shape alone — only on a non-success status.

application/json and text/event-stream are still parsed and streamed exactly as before, so no behavior is lost; only the previously fatal fallback arm changes. A POST that is still waiting on a reply keeps every existing strict check.

Tests

Added to the module's inline mod tests:

• post_notification_accepts_unusable_success_body — two rstest cases, 200 + text/plain and 200 with the Content-Type stripped. Both fail on main with UnexpectedContentType(Some("text/plain")) and UnexpectedContentType(None) respectively; both pass with the change. The second also covers the chunked-empty-body shape, which takes the identical path (no Content-Length, no Content-Type).
• post_request_still_rejects_unexpected_content_type — guard against over-loosening: a POST still awaiting a reply must keep failing. Passes before and after.

cargo clippy -p rmcp --all-features --all-targets is clean and cargo fmt --check passes. Full suite per the just test recipe: 949 passed, 0 failed. (test_with_python fails identically on pristine main in my environment because uv isn't installed — unrelated.)

[DaleSeo]

UnixSocketHttpClient::post_message_with_max_sse_event_size uses the same Content-Length and Content-Type decision tree, but its fallback still returns UnexpectedContentType for notification, response, and error POSTs. Could we apply the same awaits_no_response guard and fallback there, along with equivalent regression tests, so notifications/initialized stops failing for users of the Unix-socket transport?