fix(ingester): prevent false series limit throttling during scale-up

[yeya24]

What this PR does

During ingester scale-up, the local series limit (derived from global_limit / num_ingesters * replication_factor) shrinks immediately when new ingesters join the ring. However, existing ingesters cannot shed their series until head compaction (~2h). This causes false throttling for tenants who are within their global limit.

Root Cause

When ingesters scale from N to M (e.g., 75 → 249), the local limit drops from global / 75 * 3 = 108,000 to global / 249 * 3 = 32,530. An ingester holding 50,000 series (well under the 2.7M global limit) is immediately throttled because 50,000 > 32,530.

The series cannot redistribute until head compaction flushes idle series, so the throttling persists for up to 2 hours.

Fix

Cache the previous local limit per tenant in convertGlobalToLocalLimit(). If the global limit has not changed and the new calculated limit is lower than the cached value, return the cached (higher) limit.

The cache is only used when globalLimit == prev.globalLimit, ensuring:

• Global limit decrease: immediately enforced (cache bypassed)
• Global limit increase: recalculated fresh (cache bypassed)
• Scale-up with same limit: cached (prevents false throttle)
• Scale-down: new limit is higher, cache updates naturally

A ResetLocalLimitCache() method is provided for callers to clear the cache after head compaction, when the ingester's series count reflects its true post-resharding ownership.

How was this tested?

• Unit tests covering scale-up, scale-down, global limit changes, and cache reset scenarios
• All existing limiter tests pass without modification