tsdb: use dennwc/varint to speed up WAL decoding

[bboreham]

This is a tiny library, MIT-licensed, which unrolls the loop to go about twice as fast.

Needed to copy the unsigned-to-signed ("ZigZag") logic inline, previously provided by the binary package.

Benchmarks ("before" is including the benchmark fixes from #8645, at commit 7816674f420bc5c7505bc1eb80b54939424851d2; "after" is cherry-picking the commit from this PR onto that.):

name                                                                                old time/op    new time/op    delta
LoadWAL/batches=10,seriesPerBatch=100,samplesPerSeries=7200,exemplarsPerSeries=0-4     552ms ± 5%     490ms ± 4%  -11.16%  (p=0.008 n=5+5)
LoadWAL/batches=10,seriesPerBatch=10000,samplesPerSeries=50,exemplarsPerSeries=0-4     860ms ±21%     852ms ± 3%     ~     (p=0.548 n=5+5)
LoadWAL/batches=10,seriesPerBatch=1000,samplesPerSeries=480,exemplarsPerSeries=0-4     330ms ± 4%     278ms ± 6%  -15.53%  (p=0.008 n=5+5)

name                                                                                old alloc/op   new alloc/op   delta
LoadWAL/batches=10,seriesPerBatch=100,samplesPerSeries=7200,exemplarsPerSeries=0-4    52.1MB ± 0%    45.6MB ± 0%  -12.47%  (p=0.016 n=5+4)
LoadWAL/batches=10,seriesPerBatch=10000,samplesPerSeries=50,exemplarsPerSeries=0-4     262MB ± 3%     280MB ± 4%   +6.95%  (p=0.008 n=5+5)
LoadWAL/batches=10,seriesPerBatch=1000,samplesPerSeries=480,exemplarsPerSeries=0-4    52.1MB ± 6%    52.3MB ± 5%     ~     (p=0.548 n=5+5)

name                                                                                old allocs/op  new allocs/op  delta
LoadWAL/batches=10,seriesPerBatch=100,samplesPerSeries=7200,exemplarsPerSeries=0-4      586k ± 0%      540k ± 0%   -7.81%  (p=0.016 n=5+4)
LoadWAL/batches=10,seriesPerBatch=10000,samplesPerSeries=50,exemplarsPerSeries=0-4     2.49M ± 0%     2.49M ± 0%   +0.11%  (p=0.008 n=5+5)
LoadWAL/batches=10,seriesPerBatch=1000,samplesPerSeries=480,exemplarsPerSeries=0-4      490k ± 3%      488k ± 0%     ~     (p=0.548 n=5+5)

(there's no reason for allocs to change; just put that down to randomness in scheduling)

Looks like this should also speed up TSDB index, but I couldn't find any benchmarks relevant to that.

Timings of loading a real 5GB WAL:

before: 115.16user 16.00system 0:49.22elapsed 266%CPU (0avgtext+0avgdata 3670632maxresident)k
after:  109.99user 16.72system 0:47.00elapsed 269%CPU (0avgtext+0avgdata 3846304maxresident)k

[codesome]

Can we add a comment here saying what it is trying to do here?

[bboreham]

Added.

[codesome]

Thanks!