execution: batch by series optimization for matrix selectors with high window overlap

[afhassan]

This PR explores using a series-batching optimization to reduce peak RingBuffer memory usage for queries with long lookback windows and high cardinality.

Problem: Queries like sum(increase(metric[24h])) with a small step size create relatively large ring buffer for matrix selectors. This is done once per series and can quickly bloat memory causing OOM for high cardinality queries.

Solution: When matrix selector window overlap exceeds a certain number of steps (example 100), the optimizer switches from step batching to series batching:

• Processes all steps for a small batch of series (default: 1000)
• Reuse a fixed pool of ring buffers for all series batches
• This reduces peak memory significantly and avoids OOM

Changes:

• Added EnableHighOverlapBatching option to engine
• Extended SelectorBatchSize optimizer to detect high overlap queries
  • If detected, adjust StepsBatch to cover all steps and set a series batch size instead
• Implemented BufferPool for ring buffer reuse across series batches
• Added Clear() method to all buffer types for proper state reset before reuse

This PR is work in progress... The main blocker is validating that series batching can work for certain functions/operators that need all series present for execution

[MichaHoffmann]

im not sure this will even reach into ringbuffers, right now this only applies to vectors in direct aggregations i think?