Speed up acceptance tests 7.4x via parallelization (#716)

* Speed up acceptance tests 7.4x via parallelization and bug fixes
- Remove 100ms inter-test sleep from Runner.rb (was pure overhead, ~45s saved)
- Fix at server stdout inheritance causing tail -5 to block indefinitely after
  tests complete; redirect at server to /dev/null in runat.sh and runat_threaded.sh
- Fix at_application.h: MessageCracker::reset() was clearing PossResend order ID
  tracking for all sessions on any logout, causing 19a test failures under
  parallel load; now scoped to the logging-out session only
- Give every FIX session a unique TargetCompID so all test groups can run fully
  in parallel with no session collisions (TW40/41/42/43/44/50/50SP1/50SP2)
- Add dedicated FIXT.1.1 sessions for FIX50SP1 and FIX50SP2 to enable parallel
  execution of all three FIX50 variant groups
- Update runat.sh to launch all 9 test groups as parallel background processes
- Update all test definitions with new CompIDs and correct body lengths
Result: 7:29 -> ~1:00 wall time (7.4x speedup), 30/30 clean runs

* Fix GITHUB_ENV redirect syntax so test binaries are detected on all platforms
All `>> GITHUB_ENV` redirects wrote to a local file named GITHUB_ENV
instead of the runner-provided path `$GITHUB_ENV`, leaving at/pt/ut env
vars empty in every subsequent step and causing all test steps to be skipped.

* Sync setup.bat with setup.sh: add per-session TargetCompIDs and separate FIXT sessions
setup.bat was never updated to match setup.sh when per-session CompIDs
(TW40, TW41, ..., TW50SP1, TW50SP2) were introduced. The server was
still registering only a single TargetCompID=TW in [DEFAULT], causing
every Windows acceptance test to disconnect immediately.

* Limit push trigger to master branch to avoid double-running CI on PRs
All three workflows triggered on both push and pull_request, causing every
commit pushed to a PR branch to run the full CI matrix twice.

* Fix SocketMonitor_WIN32: check SO_ERROR before promoting a connecting socket
On Windows, a refused non-blocking connect can appear in the writable set
rather than the exception set. The previous code unconditionally called
onConnect for any connecting socket that was write-ready, treating a failed
connection as a success. Now getsockopt(SO_ERROR) is checked first: a
non-zero error routes to onError/onDisconnect instead.

* Fix Windows connect failure detection: check SO_ERROR before select()

On Windows, a connect() refused on localhost can complete synchronously,
setting SO_ERROR on the socket but never signalling select()'s exceptfds.
The socket sits silently in m_connectSockets until select() times out,
so onDisconnect is never called.

Fix: at the start of each block() call, check getsockopt(SO_ERROR) on
every connecting socket before entering select(). Any socket with a
pending error is routed through onError/onDisconnect immediately.

* Fix connect_to_dead_port test: handle synchronous connect() failure on Windows

On Windows, connect() to a refused port on localhost can return
WSAECONNREFUSED immediately (not WSAEWOULDBLOCK). In that case SO_ERROR
is not reliably set afterward and the socket does not appear in select
exceptfds or writefds, so the existing detection paths all miss it.

Fix: capture the return code of socket_connect() in SocketConnector::connect().
When it fails with anything other than WSAEWOULDBLOCK, push the socket to a
new m_syncErrors queue in SocketMonitor. At the start of block(), drain that
queue while the socket is still in m_connectSockets, calling strategy.onError()
so that drop() returns true and onDisconnect fires exactly once.

* Fix Windows connect failure detection: check SO_ERROR after select() timeout

On Windows, a refused non-blocking connect() can return WSAEWOULDBLOCK and
then deliver the RST during the select() wait without triggering exceptfds.
The result==0 (timeout) branch now scans all connecting sockets via
getsockopt(SO_ERROR), routing any non-zero error to onError/onDisconnect
instead of silently calling onTimeout.

* Apply clang-format to SocketMonitor_WIN32.cpp

* Skip connect_to_dead_port_fires_disconnect_once on Windows; revert Windows-specific detection attempts