Reduce allocations in sortKeysMapEncoder

[aaronbee]

Use one buffer for all encoded keys and values.

This significantly reduces allocations in my test of an array of 18 objects and each object has six fields.

benchmark                                old ns/op     new ns/op     delta
BenchmarkArrayJSONMarshal/jsoniter-8     67619         62885         -7.00%

benchmark                                old allocs     new allocs     delta
BenchmarkArrayJSONMarshal/jsoniter-8     500            393            -21.40%

benchmark                                old bytes     new bytes     delta
BenchmarkArrayJSONMarshal/jsoniter-8     30917         20421         -33.95%

It also brought allocations below the stdlib implementation for this test:

BenchmarkArrayJSONMarshal/stdlib-8         	   13418	     89073 ns/op	   26984 B/op	     466 allocs/op

[aaronbee]

This line is the other large source of allocations for this code. Inside this function do we have access to the length of mapIter? If so, this slice could be correctly sized when it is instantiated.

[aaronbee]

Initializing the keyValues slice to be size 8 saved another 25% in bytes allocated in my test. 8 will not always be the best value, but I think it's better than an initial size of 0 in most cases.

To get the actual length of the map, looks like reflect2 would need to expose maplen.

[codecov]

Codecov Report

Merging #424 into master will increase coverage by 0.07%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master     #424      +/-   ##
==========================================
+ Coverage   86.46%   86.54%   +0.07%     
==========================================
  Files          41       41              
  Lines        5105     5105              
==========================================
+ Hits         4414     4418       +4     
+ Misses        555      551       -4     
  Partials      136      136

Impacted Files	Coverage Δ
reflect_map.go	87.68% <100%> (ø)	⬆️
reflect_struct_decoder.go	81.89% <0%> (+0.53%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 44a7e73...f814d6c. Read the comment docs.