build(win): compile WebSocket on Windows

[EdmondDantes]

config.w32 named core, http1, formats, log, room, static and llhttp, and never src/websocket or the bundled wslay. So TrueAsync\WebSocket did not exist in a Windows build and its entire phpt group failed — most of what WINDOWS_X64_ZTS_RELEASE has been reporting as red on main. Nothing platform-specific stood in the way; the recipe simply did not list the sources.

Three things the port needed, and each is a trap worth naming:

• The wslay translation units include no PHP header. A macro that reaches ours through main/config.w32.h never arrives there, so HAVE_WSLAY and HAVE_HTTP_SERVER_WEBSOCKET are passed on the command line, exactly as fix(build): pass the compression feature macros to the Windows build #292 had to do for compression.
• wslay.h types its callbacks with ssize_t, which MSVC does not have. php-src answers this with #define ssize_t SSIZE_T in main/config.w32.h — a header those units never see. SSIZE_T itself comes from BaseTsd.h, force-included because wslay.h includes only stdint.h, stdlib.h and sys/types.h.
• The block sits below the include-path section on purpose. ADD_FLAG skips a flag that is a substring of one already added: with the wslay include path added first, the module root's own /I was dropped and every file reached through php_true_async_server.h stopped resolving.

CI needs no change — the Windows workflow does not pass --disable-all, so the new --enable-websocket (default yes, mirroring config.m4) applies on its own.

Evidence. On this Windows build the group goes from absent to 60 of 61 executed tests passing; the whole phpt tree reads 496 tests, 270 passed, 224 skipped, 2 failed.

One of those two is real and outside this change: websocket/035-recv-queue-overflow fails 2 runs of 3 here with client saw close: NULL, while the handler side reports 1013 correctly. Its client reads the socket into one buffer and looks for the CLOSE opcode at byte 0, so anything the server sends ahead of the close leaves it parsing a data frame's header forever. That is a framing assumption in the test, not a build problem, and it wants its own change.

The other failure is core/078-listener-port-zero, which belongs to #302 and is present in the working copy but not in this build.

[github-actions]

Coverage

Total lines: 82.98% → 82.99% (+0.02 pp)

File	Baseline	Current	Δ	Touched
src/core/http_connection_tls.c	73.83%	73.55%	-0.28 pp
src/http3/http3_callbacks.c	83.44%	84.18%	+0.74 pp
src/websocket/ws_session.c	90.41%	90.04%	-0.37 pp