Fix HTTP1Connection reuse crash after request outlives response

[fabianfett]

Motivation

If the response finishes before the request body stream, we use the write promise of the request's .end to tell the connection pool that the connection can be reused. This write promise can be completed synchronously inside writeAndFlush. In that case the pool may schedule the next request onto the connection before writeAndFlush returns. Afterwards we transitioned the idle read/write timeout state machines, which by then belonged to the new request. With an idle read timeout configured, this crashes with:

  HTTP1ClientChannelHandler.swift: Fatal error: Invalid state. Waiting for response data must start after request head was sent

This is a regression introduced in 1.31.0 (#879).

Changes

• Update the idle read and write timeout state machines before writing and flushing the request's .end, so they are always applied to the request that is actually ending.
• Add a Swift Testing reproducer using EmbeddedChannel that schedules a new request from onConnectionIdle while the .end write completes.
• Allow MockHTTPExecutableRequest to be created with custom RequestOptions.

Result

A request whose response finishes before its body stream no longer crashes the connection when the pool immediately reuses that connection.

[weissi]

Thanks, this looks good.

State reconciliation before callouts to arbitrary code is definitely the correct choice here.