osmpbf: direct decoding and ability to skip element types

[paulmach]

wrote https://github.com/paulmach/protoscan to provide the low-level ability to walk through protobuf data. This allows for:

1. fast decoding with lower memory usage since the intermediate of the decoding proto objects can be skipped. This resulted in 45% less allocs and 20% less memory usage on the decoding benchmark.
2. ability to skip certain types by moving past the encoded proto data directly. This skips all the work of decoding and building object that are not needed. However, it still requires the blocks be decompressed, so there still is some work.

I feel pretty good about this change, however writing code using protoscan is similar to handwriting the autogenerated protobuf code which could cause errors.