feat: compress kept trace decision message

[VinozzZ]

Which problem is this PR solving?

Due to the multiplication effect of pubsub messages through Redis, we would like to bring down the amount of network traffic sent by Redis.

One of the biggest contributors is TraceDecison message used to communicate kept trace decisions.

This PR uses snappy to compress the data and then encoded using gob

Here's the benchmark results comparing to the original JSON encoding.
In the BenchmarkCompressionSizes test, I generated a batch of 1000 trace decisions as the input. The amount of decrease in data size is 90%

goos: darwin
goarch: arm64
pkg: github.com/honeycombio/refinery/collect
cpu: Apple M2 Max
BenchmarkDynamicJSONEncoding-12          	   3139	   353164 ns/op	 443032 B/op	      3 allocs/op
BenchmarkDynamicCompressedEncoding-12    	   3050	   383788 ns/op	 475333 B/op	     62 allocs/op
BenchmarkDynamicJSONDecoding-12          	    561	  2114997 ns/op	 590167 B/op	   5018 allocs/op
BenchmarkDynamicCompressedDecoding-12    	   3285	   360892 ns/op	 445667 B/op	   5244 allocs/op
BenchmarkCompressionSizes/JSON_Encoding-12         	   5000	   238046 ns/op	 164091 B/op	      2 allocs/op
--- BENCH: BenchmarkCompressionSizes/JSON_Encoding-12
    trace_decision_test.go:305: JSON Encoding: Total Batch: 1, Total Size: 160671 bytes, Average Size: 160671.00 bytes
    trace_decision_test.go:305: JSON Encoding: Total Batch: 100, Total Size: 16227771 bytes, Average Size: 162277.71 bytes
    trace_decision_test.go:305: JSON Encoding: Total Batch: 5000, Total Size: 819582771 bytes, Average Size: 163916.55 bytes
BenchmarkCompressionSizes/Snappy_Compression-12    	   5052	   236998 ns/op	 299547 B/op	     59 allocs/op
--- BENCH: BenchmarkCompressionSizes/Snappy_Compression-12
    trace_decision_test.go:319: Snappy Compression: Total Batch: 1, Total Size: 12115 bytes, Average Size: 12115.00 bytes
    trace_decision_test.go:319: Snappy Compression: Total Batch: 100, Total Size: 1223615 bytes, Average Size: 12236.15 bytes
    trace_decision_test.go:319: Snappy Compression: Total Batch: 5052, Total Size: 62428595 bytes, Average Size: 12357.20 bytes

I also used sync.Pool for both compression and decompression buffers as well as snappy.Writer to reduce allocation

[kentquirk]

This looks good! I had a couple of comments, but they're optional.

[kentquirk]

I think this is the most efficient way to do that:

Suggested change

	out := make([]byte, buf.Len())
	copy(out, buf.Bytes())
	return out, nil
	return bytes.Clone(buf.Bytes()), nil

[kentquirk]

Why not also compress the dropped messages? Even though they're unlikely to do much better than 50% compression (they're essentially hex-coded noise most of the time), that's not nothing given how many we send.