Reduce fmt.Sprintf allocations in query encoding

[kgeckhart]

This PR eliminates some calls to fmt.Sprintf in favor of string concatenation when the values being concatenated are strings. This should reduce the memory required to construct queries with complex objects / larger arrays. Arrays which are not flat should see a further improvement as the prefix is now computed once ahead of time instead of doing it with an fmt.Sprintf call on every element.

Included are some benchmarks which I used to decide that the maps + array use cases with fmt.Sprintf seem okay. Dropping fmt.Sprintf is faster but won't lead to any reduction in allocations and in the case of arrays a size between 100 and 499 is actually more allocations.

[Madrigal]

Thanks for your contribution. Just letting you know this is still on our radar, but we've been a bit busy.

For context, was there a particular scenario where this was causing a performance issue for you?

[kgeckhart]

Thanks for the response. I work on an OSS CloudWatch exporter https://github.com/prometheus-community/yet-another-cloudwatch-exporter/. When running high volumes we see some rather spiky memory and I was trying to reduce it a bit. This shows the top allocators from pprof after running for some time,

fmt.Sprintf was a lot higher than I expected it be on the list.

Almost all of the fmt.Sprintf allocations appear to stem from these two functions

when serializing the large GetMetricDataInput payload.