[Trombo38]

EXTERNAL_SERVICE on Windows (local process units)

What

EXTERNAL_SERVICE on an External API preset now works on Windows, not only Linux.
Today the four platform hooks (unitAction, unitActiveState, unitPIDs,
waitGPUsReleased) are stubs on Windows that return
"EXTERNAL_SERVICE n'est pris en charge que sous Linux" — so a preset that names a
service fails to switch on Windows instead of starting its engine.

This PR implements them on Windows with the same contract the Linux systemd units
already provide:

• switch to the preset → start the named service
• switch away → stop it, then wait for the GPUs to be released
  (nvidia-smi --query-compute-apps=pid, ported from sys_service_linux.go)
  before the next engine loads — so one engine owns the GPUs at a time.

On Windows the "unit" is a local process defined by a JSON file in
$AJEAN_HOME\services\ajean-<name>.json (the new services/ folder, created on
first use — it does not exist in a current install):

{
  "command": "I:\\Strata\\.venv\\Scripts\\python.exe",
  "args": ["I:\\Strata\\serve\\server.py", "--engine", "strata",
           "--config", "I:\\Strata\\strata-swift-iq2_xs.json", "--port", "8080"],
  "cwd": "I:\\Strata",
  "env": { "PYTHONUTF8": "1" }
}

command/cwd may be relative (resolved from the file's folder); env is merged
over the inherited environment. The process is launched detached (same flags as
ajean serve), tracked by a PID file, logged to $AJEAN_HOME\<unit>.log, stopped
with taskkill /T /F — the exact primitives the local engine supervisor already
uses, no admin rights, no external tools.

The preset itself is unchanged and portable across platforms:
EXTERNAL_SERVICE=ajean-strata-iq2xs (same ^ajean-[a-z0-9][a-z0-9-]{0,40}$ regex).

Why the GPU-mutex is safe

SwitchToPreset (backend_presets.go) already orders the calls correctly and this
PR does not change that path — it only makes the Windows hooks real:

• → external preset: serviceAction("restart") first (which on Windows +
  external config stops the local llama-server and waits for the port to free),
  then syncExternalService starts the external service.
• → local preset: syncExternalService stops the external unit,
  waitGPUsReleased waits for its PIDs to leave the GPU, then llama.cpp restarts.

So the two engines never coexist — the Windows equivalent of the Linux
Conflicts=ajean-engine.service, obtained by call order.

Files

• internal/ajean/run.go — servicesDir() = $AJEAN_HOME\services
• internal/ajean/sys_service_windows.go — remove the 4 "Linux only" stubs
• internal/ajean/sys_service_windows_extservice.go (new) — the 4 hooks
• internal/ajean/sys_service_windows_extservice_test.go (new) — def parsing,
  state tracking, GPU-mutex order (via syncExternalService)
• internal/ajean/backend_extservice.go — doc comment updated (no behavior change)
• README.md / README.fr.md — "Third-party engine" section now Linux & Windows

Testing (CI gates)

• gofmt -l . (Go 1.25) — clean.
• go vet ./... — clean.
• staticcheck@latest — 0 warnings (project is zero-warning).
• Cross-compile CGO_ENABLED=0 — 4/4: linux/amd64, linux/arm64, windows/amd64, windows/arm64.
• go test ./... (with AJEAN_HOME isolated) — all pass, except TestMoeMigrate
  which fails on this Windows host due to a symlink privilege it does not have
  ("A required privilege is not held by the client"); it passes on the CI ubuntu
  runner and is unrelated to this change (the PR touches no MoE or symlink code).
• All EXTERNAL_SERVICE / external tests pass: TestSyncExternalService,
  TestSyncExternalServiceRejectsForeignUnits, TestExternalServiceReady,
  TestExternalPresetEditKeepsServiceAndMachine, plus the new
  TestExtServiceDefParsing, TestExtUnitStateTracking,
  TestWindowsExternalServiceMutex.
• Manual end-to-end (run, then removed): a dummy engine defined in
  services/ajean-e2e.json was started via syncExternalService, reached
  externalServiceReady (unit active + /health 200), and was stopped cleanly
  (port no longer listening).