fix(web): fall back to a free port when the configured one is taken

[lenucksi]

Problem

The plugin crashes when its configured web-server port (default 4200) is already in use:

• PTYServer.startWebServer calls Bun.serve({ port, ... }) with no EADDRINUSE handling.
• The V2 setup autostarts the server with a bare await getOrCreateServer(...) — a busy port rejects PTYServer.createServer, which makes plugin setup throw, i.e. the plugin dies at load.
• There is neither a fallback to the next free port nor a check for an already-running opencode-pty server on that port.

Reproduced: with one opencode instance already serving on 4200, starting a second instance with the default config crashes its plugin.

Fix

• Port fallback: keep port: 0/unset semantics (OS-assigned, never conflicts). For an explicit port, try it and on EADDRINUSE walk port+1 … port+10, with port 0 as the final fallback. Every attempt logs which port is busy; the final port is logged too. server.url/getWsUrl are unchanged, so the dynamic port flows into the web UI and /pty-show-server-url correctly.
• EADDRINUSE detection: checks the Bun error code, with a message-pattern fallback for differently-shaped environments.
• Existing-server probe (diagnostic only): before binding an explicit port, a short /health probe detects an already-running opencode-pty server and logs that clearly instead of crashing. The bind loop stays the TOCTOU-safe authority. Cross-process "server reuse" is intentionally not attempted: the PTY manager is in-process per plugin instance, so adopting a foreign server would show a foreign session list.
• Crash-proof V2 setup: the autostart is wrapped in try/catch — on failure it warns and continues; the PTY tools stay fully functional in-process, and later on-demand server invocations get the fallback automatically.

Tests

15 new tests in test/port-fallback.test.ts:

• configured port taken → resolves, binds the next free port, /health OK on the new port
• configured port + all 10 fallbacks blocked → OS-assigned port, still healthy
• port: 0 → random OS-assigned port
• V2 Plugin.setup with autostart: true and a busy port → resolves, warning logged, getActiveServer() populated
• pure helpers (resolution, candidate generation incl. 65535 cap, EADDRINUSE detection)

Verification

• bun run format / lint / typecheck / build:prod — clean
• bun test — 97 pass; the only failures are the same 4 that also fail on main in this environment (2 package tests need npm, which is absent here; 2 are timing-sensitive and pass in CI)
• Live check on this host: second instance with PTY_WEB_PORT=4200 while 4200 was held → bound to 4201, healthy, original server untouched, log shows the "already runs / next port" warnings

Standalone PR (based on main, 2 commits) — intentionally not part of #61.