db: fine-grained profiling

[RaduBerinde]

This issue tracks implementing Pebble infrastructure for obtaining data on resource utilization (like CPU, disk bandwidth) broken down on various axes:

• CRDB key type (lock, txn, intent, table)
• read path / write path
• compression / decompression
• flushes / compaction (by target level)

Possibilities for tracked resources:

• CPU usage
• disk write bandwidth, iops
• disk read bandwidth, iops
• bloom filter negatives, false positives

The data should allow easily answering questions like "in this workload, how much CPU usage was caused by compactions of intent keys?"

We can implement this using pprof labels. For resources other than CPU we can maintain the statistics ourselves, as a map from label subsets. We could output them as pprof files or some other format.

This infrastructure can be behind a build flag to avoid adding overhead.

Jira issue: PEBBLE-260