Use buffer pools for decompression

[ronanh]

While profiling my application, I've noticed that Sarama spends a significant amount of time growing slices when it's decompressing messages:

This patch uses an intermediate buffer from a pool for decompression, to avoid these allocations.
When the final size is know, the result buffer can be allocated and copied.

After the patch:

[ronanh]

I have signed the CLA!

[hindessm]

Looks very good to me:

before:
BenchmarkDecompression/lz4_2048_compressable-4          578866    9368 ns/op    8533 B/op   11 allocs/op
BenchmarkDecompression/lz4_2048_lessCompressable-4      830926    6521 ns/op    8506 B/op   11 allocs/op
BenchmarkDecompression/lz4_2048_compressable-16         804277    7391 ns/op    8544 B/op   11 allocs/op
BenchmarkDecompression/lz4_2048_lessCompressable-16    1221556    4904 ns/op    8538 B/op   11 allocs/op

after:
BenchmarkDecompression/lz4_2048_compressable-4          792148    7512 ns/op    2373 B/op    7 allocs/op
BenchmarkDecompression/lz4_2048_lessCompressable-4     1302768    4614 ns/op    2395 B/op    7 allocs/op
BenchmarkDecompression/lz4_2048_compressable-16         920294    5854 ns/op    2364 B/op    7 allocs/op
BenchmarkDecompression/lz4_2048_lessCompressable-16    1544532    3909 ns/op    2420 B/op    7 allocs/op