fix(e2e/discovery): Let a crash-looping fixture fail instead of restarting forever (#54350)

### What does this PR do?

Stops the fixture services restarting forever when they are broken, so a fixture that cannot start reports itself as a failed unit.

The units set `StartLimitIntervalSec=0`, which disables systemd's start rate limit entirely. Combined with `Restart=always` and `RestartSec=1`, a service that fails immediately restarts forever: it stays in `activating (auto-restart)` and never reaches `failed`.

This PR enables the limit (`StartLimitIntervalSec=60`, `StartLimitBurst=5`) so a service that keeps failing gives up and the unit ends in `failed`.

### Motivation

DSCVR-621. There are some cases were discovery is failing but in some cases we can't clearly distinguish discovery really failing from the server just not starting up so make the service startup fail explicitly.

### Describe how you validated your changes

In a container running systemd 249 (the same version as the test image), with a service that exits after 450ms to match the observed rails failure:

| `[Unit]` config | State after 25s | `NRestarts` |
|---|---|---|
| `StartLimitIntervalSec=0` (before) | `activating (auto-restart)` | **16, still climbing** |
| `StartLimitIntervalSec=60` + `StartLimitBurst=5` (after) | **`failed`** (`Result: exit-code`) | 5 |

### Additional Notes

One deliberate trade-off: a fixture that currently succeeds only after more than five automatic restarts would now fail instead. That is right outcome — a fixture needing six attempts is broken and should say so.


Co-authored-by: vincent.whitchurch <vincent.whitchurch@datadoghq.com>

Author: vitkyrka

test/new-e2e/tests/discovery/testdata/provision/provision.sh

@@ -33,11 +33,20 @@ install_systemd_unit() {
     done <<< "${extraenv} "
   fi
 
+  # The start limit is left enabled on purpose. These services are
+  # Restart=always with RestartSec=1, so one which fails immediately restarts
+  # forever: it stays in "activating (auto-restart)" and never reaches "failed",
+  # which makes a broken fixture look like a service the Agent failed to
+  # discover rather than a service that is not running. Letting the limit stop
+  # it turns that into a failed unit instead. The counter is reset by an
+  # explicit systemctl stop, so the start/stop cycles the tests do between
+  # subtests never accumulate towards it.
   cat > "/etc/systemd/system/${name}.service" <<- EOM
 [Unit]
 Description=${name}
 After=network.target
-StartLimitIntervalSec=0
+StartLimitIntervalSec=60
+StartLimitBurst=5
 
 [Service]
 Type=simple