[metricbeat] system/filesystem: prefer hostfs-prefixed mount points when deduplicating devices

Nobody has claimed this yet.

When a block device is mounted multiple times (e.g. bind mounts), the `system/filesystem` metricset deduplicates entries keeping only one mount point per device. The current logic always picks the **shortest** path. This is incorrect when `hostfs` is configured (the standard container-based monitoring setup), because container-local paths like `/` are shorter than the host paths like `/hostfs`, so the host mount points — the ones the user actually wants — get discarded.

The deduplication should use the following precedence for choosing which mount point to keep for a given device:
1. Prefer paths under the `hostfs` prefix when `hostfs` is configured.
2. Among paths with equal prefix status, prefer the shorter path (current behavior, as tiebreaker).

When `hostfs` is not configured, behavior stays the same.

The fix spans two repositories:
- **`elastic-agent-system-metrics`** — the [`filterDuplicates`](https://github.com/elastic/elastic-agent-system-metrics/blob/main/metric/system/filesystem/filesystem.go#L150-L174) function needs to accept the `resolve.Resolver` (or the hostfs prefix) and apply the new precedence logic.
- **`elastic/beats`** — the [`Fetch` method in `metricbeat/module/system/filesystem/filesystem.go`](https://github.com/elastic/beats/blob/main/metricbeat/module/system/filesystem/filesystem.go#L80-L107) calls `GetFilesystems`, which calls `filterDuplicates`. After the library is updated, bump the dependency and add a test in metricbeat that proves hostfs-prefixed paths are preferred.

Start with filterDuplicates in elastic-agent-system-metrics/metric/system/filesystem/filesystem.go and trace how Fetch in metricbeat/module/system/filesystem/filesystem.go reaches it through GetFilesystems. Check the existing filesystem tests, then verify that hostfs-prefixed paths win when configured, shorter paths remain the tiebreaker, and behavior is unchanged without hostfs.

Written by the indexing model from the issue text.