A number of improvement to improve downsampling performance

[martijnvg]

Number of changes:

• Improve downsample dimension processing (Avoid reading unnecessary dimension values when downsampling #124451), which avoids unnecessary reads.
• Make downsampling faster by buffering docids and bulk processing (Improve downsample performance by buffering docids and do bulk processing. #124477), which share some overhead across a bulk of doc ids instead of for each docid.
• Change MetricFieldProducer to use SortedNumericDoubleValues instead of FormattedDocValues, which saves unneeded number conversations and casts.
• Remove unneeded Metric abstraction in MetricFieldProducer, which speeds up rolling up gauges/counters.

All changes result into lowering downsample latency:

• For the tsdb downsampling benchmark track, the 1 minute interval downsample operation latency dropped by ~30%
• For the tsdb downsampling benchmark track, the 1 hour interval downsample operation latency dropped by ~50%