[prototype] Add type-specific decoder code generator

[tsholmes]

This adds a new objconvgen entrypoint that generates a optimized decoder for a specific type. Since the decoder doesn't need to use reflection and can return all results by-value, it attains a significant speedup from the generic parsing.

Since this is a prototype, I'd like to get some feedback on the approach I took. Any and all suggestions and feedback are welcome. This is a pretty big feature, so I'd like everyone to be on board with it.

Sample benchmark on a structure with strings, times, and nested arrays

BenchmarkDecodeCBOR-8                  200000          6431 ns/op          80 B/op           6 allocs/op
BenchmarkDecodeJSON-8                  200000          8607 ns/op          80 B/op           6 allocs/op
BenchmarkDecodeCBORGenerated-8         500000          3319 ns/op          48 B/op           5 allocs/op
BenchmarkDecodeJSONGenerated-8         300000          5425 ns/op          48 B/op           5 allocs/op

What's supported right now:

• structures
• arrays (not slices)
• primitives
• time.Time
• aliased types
• types in different packages

What's not supported right now:

• slices
• maps
• time.Duration
• []byte/[N]byte special case

What definitely needs to be done before this could be merged:

• support json and objconv struct tags
• write the output to a file and goimports/gofmt it
• add tests on some generated code
• add benchmarks on some generated code

What would be nice to clean up before this is merged:

• remove dependence of generated code on github.com/segmentio/objconv/generate/util package