perf(arrow/array): avoid temporary buffers when concatenating

[fallintoplace]

Summary

• copy fixed-width values directly from source buffer ranges into the output buffer
• avoid creating one temporary memory.Buffer for every input chunk
• add a benchmark with a fixed number of values and increasing chunk counts

Benchmarks

Apple M1 Pro, 65,536 int64 values:

chunks	old time	new time	delta	old allocs	new allocs
1	65.92 us	59.86 us	~	7	5
8	48.58 us	51.12 us	~	15	6
64	64.33 us	64.18 us	~	71	6
1,024	154.88 us	97.89 us	-36.80%	1,031	6
8,192	571.7 us	214.5 us	-62.48%	8,199	6

At 1,024 chunks, B/op drops by 14.22%. At 8,192 chunks, it drops by 52.06%.

Tests

• go test ./...