[BUG] MetaDataCache: unbounded CACHE (maximumSize=Integer.MAX_VALUE) + MAPPING Set never shrunk

## Description
`CACHE` is a `WindowTinyLFUMap` built with `maximumSize = Integer.MAX_VALUE`, which disables size-based eviction entirely; there is no `expireAfterWrite`. The class comment at line 148 even acknowledges OOM risk yet configures an unbounded size. `MAPPING` (`ConcurrentMap>`) is populated by `initCache`: for wildcard metadata paths the reverse-mapping `ConcurrentSkipListSet` under each `metaPath` collects every distinct request path ever matched. `clean(key)` removes CACHE entries but **never clears the Set itself and never removes it from MAPPING**; `remove(MetaData)` only calls `clean(path)`.

## Location
```
shenyu-plugin-base/.../cache/MetaDataCache.java:48 (CACHE), 53 (MAPPING), 99-107 (clean), 147-157 (initCache)
```

## Impact
`MetaDataCache.obtain(path)` is called every request from the global plugin and divide/context-path/rewrite plugins. High-cardinality request paths (e.g. `/api/users/{userId}` with millions of distinct IDs) cause monotonic, unbounded heap growth in both CACHE and the MAPPING `ConcurrentSkipListSet` (memory-heavy per entry) that survives metadata refreshes. Slow OOM over days/weeks for public-facing REST gateways.

## Suggested fix
Set a real configurable `maximumSize` and `expireAfterAccess`; in `clean(key)`, after removing CACHE entries, also `MAPPING.get(key).clear()` and `MAPPING.remove(key)`.

## Related existing issue(s)
#6506 is the CachePlugin (response caching); this is the metadata path cache. #6479 is discovery-upstream cache staleness; this is selector/metadata cache growth.

_Identified during the 2026-08-02 audit; full list in [`docs/issue-candidates-2026-08-02.md`](docs/issue-candidates-2026-08-02.md)._

Start in shenyu-plugin-base/.../cache/MetaDataCache.java at CACHE, MAPPING, clean, remove, and initCache. Trace obtain and metadata refresh behavior, then verify that cache entries have bounded lifetime and that reverse-mapping sets and keys are removed during cleanup without retaining request paths.

Written by the indexing model from the issue text.