Optimization: Avoid excessive decoding and allocation in exponential histogram last/first_over_time

[JonasKunz]

Follow up issue for #138639 (comment).

For the implementation of first_over_time and last_over_time, we need to keep track of the currently "best" exponential_histogram. This is at the moment done by naively copying the histogram, causing both an allocation and a decoding of the histogram stored in the block.

We should optimize this by introducing something similar to a BreakingBytesRefBuilder, e.g. a resuable EncodedHistogramHolder:

• It should internally use a BreakingBytesRefBuilder to store the histogram bytes and be reusable
• It should be populateable from an ExponentialHistogram returned from a block without decoding: It should "see through" e.g. the CompressedHistogram´ and directly access and copy the underlying bytes[]`
• It should be addable to an ExponentialHistogramBlockBuilder without requiring an encoding: The data should be directly copied instead

[elasticsearchmachine]

Pinging @elastic/es-analytical-engine (Team:Analytics)