Fixes cockroachdb#62081. It had rotted after we removed the trace.mode setting. It
now makes use of a debug-only cluster setting that if enabled, installs
real spans all sql statements. We'll defer the actual investigation into
the cause of the slow-down in future PRs.

    $ make bench PKG=./pkg/bench  TESTFLAGS='-benchtime=10000x -count 20' \
        BENCHES='BenchmarkTracing' BENCHTIMEOUT=50m > benchmark-tracing.txt
    $ cat benchmark-tracing.txt | grep -v tracing=t | sed 's/tracing=f//' > old.txt
    $ cat benchmark-tracing.txt | grep -v tracing=f | sed 's/tracing=t//' > new.txt
    $ benchstat old.txt new.txt

    name                                   old time/op    new time/op    delta
    Tracing/Cockroach//Scan1-24               382µs ± 2%     412µs ± 8%   +7.68%  (p=0.000 n=10+10)
    Tracing/Cockroach//Insert-24              579µs ± 2%     609µs ± 6%   +5.10%  (p=0.000 n=10+10)
    Tracing/MultinodeCockroach//Scan1-24      799µs ± 2%     885µs ± 1%  +10.76%  (p=0.000 n=10+9)
    Tracing/MultinodeCockroach//Insert-24    1.09ms ± 1%    1.14ms ± 2%   +5.04%  (p=0.000 n=9+10)

    name                                   old alloc/op   new alloc/op   delta
    Tracing/Cockroach//Scan1-24              25.3kB ± 5%    32.0kB ± 4%  +26.55%  (p=0.000 n=10+10)
    Tracing/Cockroach//Insert-24             38.0kB ± 6%    42.2kB ± 5%  +11.02%  (p=0.000 n=10+10)
    Tracing/MultinodeCockroach//Scan1-24     77.7kB ± 8%    97.1kB ±12%  +25.05%  (p=0.000 n=10+10)
    Tracing/MultinodeCockroach//Insert-24     107kB ± 8%     115kB ± 7%   +7.66%  (p=0.023 n=10+10)

    name                                   old allocs/op  new allocs/op  delta
    Tracing/Cockroach//Scan1-24                 256 ± 1%       280 ± 1%   +9.10%  (p=0.000 n=9+9)
    Tracing/Cockroach//Insert-24                301 ± 2%       321 ± 1%   +6.64%  (p=0.000 n=10+10)
    Tracing/MultinodeCockroach//Scan1-24        787 ± 2%       921 ± 2%  +17.04%  (p=0.000 n=9+9)
    Tracing/MultinodeCockroach//Insert-24       748 ± 1%       805 ± 2%   +7.61%  (p=0.000 n=9+9)

Release note: None

Instead of the incoming ctx. Going through the ctx package is expensive,
and log tags are unlikely to change between parent and child spans.

---

Actually, drop log tags altogether.

Release note: None

Release note: None

- Reduce lock contention by only deleting root spans from registry.
- Don't materialize span recordings unless absolutely needed.
- Compute span duration outside of mutex

Release note: None

Tracer.AlwaysTrace is redundant. We can also avoid doing opentracing
stuff if we have no external sinks configured.

Release note: None

Release note: None

This reverts commit 1631298.

Release note: None

- use an optimizedCtx for collapsed spans
- avoid stepping through interface for carrier
- mutate incoming ctx's metadata.MD directly
- disable emptry registry check, we broke it a few commits ago
([wip,dnm] tracing: move from interface type; avoid string.ToLower)
- skip some common span tags

Release note: None