db: improve context propagation for tracing of slow block loads

[sumeerbhola]

We sometimes see traces in CockroachDB which suggest slowness in Pebble, but the InternalIteratorStats printed don't indicate block cache misses. This is possibly because those stats are incomplete (don't include Reader creation or loading of various block types). We have a LoggerAndTracer passed to Pebble, and tracing of slow block loads (5ms is the slow threshold in Reader.readBlock), since #2055. But there are some significant gaps:

• Many Reader.readBlock calls are using a background context, e.g. Reader.readRangeDel, readRangeKey, readMetaindex. These should use a real context.
• tableCacheValue.load which creates a new Reader should accept a context. We have one in the callpath that originates in tableCacheShard.newIters.

Jira issue: PEBBLE-36

[Radu] Making a checklist for context propagation:

• tableCacheValue.load (in the newIters path)
• Ingest methods
• Value block load
• Reader block reads (readRangeDel, readRangeKey, readMetaindex)

[RaduBerinde]

Removing the P-2 label because the important paths have been addressed. There is only one path around value blocks that is more tricky.