xdr and horizon: Optimize compress-marshaling of ledger keys

[2opremio]

15x CPU time improvement, all allocations were removed

Before:

goos: darwin
goarch: amd64
pkg: github.com/stellar/go/benchmarks
cpu: Intel(R) Core(TM) i7-1068NG7 CPU @ 2.30GHz
BenchmarkXDRMarshalCompress
BenchmarkXDRMarshalCompress-8   	  608031	      1970 ns/op	    2048 B/op	      59 allocs/op
PASS

After:

goos: darwin
goarch: amd64
pkg: github.com/stellar/go/benchmarks
cpu: Intel(R) Core(TM) i7-1068NG7 CPU @ 2.30GHz
BenchmarkXDRMarshalCompress
BenchmarkXDRMarshalCompress-8   	 8785383	       133.8 ns/op	       0 B/op	       0 allocs/op
PASS

One of the remaning allocations is, in fact, caused by a call to PoolID.Encode which I think can be removed.

[2opremio]

When writing this PR I started to wonder why don't we simply marshal the keys normally and hash them?

This would:

1. Save a lot more space
2. Save us from maintaining this compression hack

In other words, do we really need this sort of compression if the result is opaque?

[2opremio]

@bartekn PTAL

[2opremio]

I am surprised this casting to []byte from string didn't cause an allocation. The escape analysis is probably getting better in Go (it wasn't nearly as good as I would had liked to)