Fix 100% CPU spin when proxy closes connection during CONNECT

[LukeButters]

Background

Fixes: LEV-2050

A polling Tentacle configured with an HTTP proxy can get stuck at 100% CPU until the service is restarted (LEV-2050).

HttpProxyClient.SendConnectionCommandAsync reads the proxy's CONNECT response one byte at a time until it sees \r\n\r\n. It never checked whether ReadAsync returned 0, which means the connection was closed. So when a proxy accepted the TCP connection and then closed it before finishing its response, for example while restarting, the loop spun forever. Each pass also called StringBuilder.ToString(), so it allocated continuously and kept the GC busy too. The read timeout doesn't help, because a read on a closed connection returns 0 straight away instead of waiting. The only way out was cancellation, which in practice meant stopping the service.

The loop was introduced in #468.

Change

• If ReadAsync returns 0, throw a ProxyException (causedByNetworkError: true). The existing retry and polling backoff then handle it as usual.

Tests

• New HttpProxyClientFixture with a fake proxy that accepts the connection, reads the CONNECT request, then closes. One case closes without sending anything; the other sends a partial response first.
  • Before the fix, both cases spun until the 10s cancellation, using about 92–97s of CPU.
  • After the fix, both fail promptly with a ProxyException.
• ProxyFixture and the new fixture pass on net8.0 (macOS): 52 tests.

Not covered

• I have not run this on Windows or net48. The customer was on net48 on Windows.
• When proxy connection setup fails, the TcpClient is never disposed, so each failed attempt leaks a socket. That's separate from the CPU problem and is left for a follow-up.