Reuse key buffers during hashing

[StephenButtolph]

Why this should be merged

This removes the last memory allocation inside of HashChangedNodes (except for the allocation performed for tracing.

Before:

Benchmark_HashChangedNodes/1-12  		 3299274	       361.6 ns/op	     144 B/op	       3 allocs/op
Benchmark_HashChangedNodes/10-12 		  407604	      2992   ns/op	     161 B/op	       5 allocs/op
Benchmark_HashChangedNodes/100-12         	   39985	     30044   ns/op	     280 B/op	      19 allocs/op
Benchmark_HashChangedNodes/1000-12        	    3674	    332787   ns/op	    2746 B/op	     293 allocs/op
Benchmark_HashChangedNodes/10000-12       	     356	   3489978   ns/op	   14262 B/op	    1573 allocs/op
Benchmark_HashChangedNodes/100000-12      	      19	  59557794   ns/op	  355809 B/op	   37157 allocs/op

After:

Benchmark_HashChangedNodes/1-12  		 3664717	       337.9 ns/op	     144 B/op	       3 allocs/op
Benchmark_HashChangedNodes/10-12 		  438346	      2842   ns/op	     160 B/op	       4 allocs/op
Benchmark_HashChangedNodes/100-12         	   42628	     28271   ns/op	     272 B/op	      11 allocs/op
Benchmark_HashChangedNodes/1000-12        	    3680	    321290   ns/op	    2464 B/op	     148 allocs/op
Benchmark_HashChangedNodes/10000-12       	     369	   3150519   ns/op	   12704 B/op	     788 allocs/op
Benchmark_HashChangedNodes/100000-12      	      22	  52984970   ns/op	  297376 B/op	   18580 allocs/op

Honestly it's getting pretty hard to benchmark these changes... The variability of the benchmark is higher than the improvements. This seems to give as much as another 12% reduction in runtime... But I'm not especially confident in that number. I am however confident that after this PR and #2901 are merged we'll see:

Benchmark_HashChangedNodes/1-12  		 3578389	       333.3 ns/op	     144 B/op	       3 allocs/op
Benchmark_HashChangedNodes/10-12 		  432597	      2875   ns/op	     144 B/op	       3 allocs/op
Benchmark_HashChangedNodes/100-12         	   41547	     28324   ns/op	     144 B/op	       3 allocs/op
Benchmark_HashChangedNodes/1000-12        	    3766	    315036   ns/op	     144 B/op	       3 allocs/op
Benchmark_HashChangedNodes/10000-12       	     345	   3263585   ns/op	     144 B/op	       3 allocs/op
Benchmark_HashChangedNodes/100000-12      	      24	  51812141   ns/op	     144 B/op	       3 allocs/op

How this works

• Maintains a persistent set of keyBuffers such that HashChangedNodes only performs allocations during the first few calls.
• Replaces the semaphore with a byte pool that implements the buffer sharing.
• Shares the keyBuffer with recursive calls to hashChangedNode.
• Avoids copying n.key.value on each call to hashChangedNode.

How this was tested

• CI
• Added benchmarks around the byte pool

[danlaine]

just nits

[darioush]

I added some comments about the comments.