Make intern.Table completely lock-free

[mcy]

This change removes all locks in intern.Table, replacing the map with a sync.Map and the table with an atomicx.Log. This eliminates virtually all lock contention, resulting in wins on all workloads.

The sync.Map is used to ensure that elements are only ever allocated once, using LoadOrStore to select the goroutine responsible for selecting the intern ID for a string. This ID is selected locklessly using atomicx.Log. This makes every operation, except for rare atomicx.Log resizes, lockless.

Here is the before and after benchmarks for a mix of hit-heavy and miss-heavy workloads (Npct means that N% of intern calls are with an uninterned value).

BenchmarkIntern/0pct-10                      361           2840396 ns/op
BenchmarkIntern/10pct-10                     332           3720667 ns/op
BenchmarkIntern/50pct-10                     225           5171647 ns/op
BenchmarkIntern/100pct-10                    230           5382226 ns/op

BenchmarkIntern/0pct-10                     6547            160613 ns/op
BenchmarkIntern/10pct-10                    1818            651644 ns/op
BenchmarkIntern/50pct-10                    1003           1279426 ns/op
BenchmarkIntern/100pct-10                    696           2215831 ns/op

Notice that we see a 17x improvement for a hit-heavy workload, a 4x improvement for the 50% workload. This patch is an unambiguous performance improvement.

To validate this finky change, in addition to the compiler's tests, I've written a test that hammers an intern table directly by simulating a thundering herd.

[mcy]

Ok so I finally go around to doing a macrobenchmark (compiling all of googleapis with sourcecodeinfo output) with this change, and it's about a 20% improvement, so I'm going to merge it.