Concurrent decoder

[AlekSi]

I implemented concurrent decoder of OSMData blocks. Also I added some small optimizations.

Benchmark results for russia-european-part-latest.osm.pbf. master:

$ go test -v -bench=. -cpu 1,2,4
BenchmarkDecoder           1    91509929417 ns/op
    decode_test.go:154: Done in 91.510 seconds. Total: 77561149, Nodes: 70272938, Ways: 7082532, Relations: 205679
BenchmarkDecoder-2         1    86301976850 ns/op
    decode_test.go:154: Done in 86.302 seconds. Total: 77561149, Nodes: 70272938, Ways: 7082532, Relations: 205679
BenchmarkDecoder-4         1    82508371534 ns/op
    decode_test.go:154: Done in 82.508 seconds. Total: 77561149, Nodes: 70272938, Ways: 7082532, Relations: 205679

this branch:

$ go test -v -bench=. -cpu 1,2,4
BenchmarkDecoder           1    85276857960 ns/op
    decode_test.go:199: Done in 85.277 seconds. Total: 77561149, Nodes: 70272938, Ways: 7082532, Relations: 205679
BenchmarkDecoder-2         1    56638467854 ns/op
    decode_test.go:199: Done in 56.638 seconds. Total: 77561149, Nodes: 70272938, Ways: 7082532, Relations: 205679
BenchmarkDecoder-4         1    42139943615 ns/op
    decode_test.go:199: Done in 42.140 seconds. Total: 77561149, Nodes: 70272938, Ways: 7082532, Relations: 205679

Objects are returned from Decode() in the same order as before (as they present in file).

Incompatible changes:

• I added Start() method to Decoder. While it can be lazily called from Decode(), I think explicit call is better.
• Error handling is simplified. Now decoder expects exactly one OSMHeader block followed by OSMData blocks. Other block types or other order is considered an error. I'm not aware about other types or order in the wild, so I think it's practical.

[qedus]

I have a couple of comments but think the speed improvement and architecture are great.