internal/compile: optimize the encoding (google#38)

The simple gob encoding was showing up in profiles.
This change replaces it with a simple hand-written binary encoding
using strings and varints.  Strings and byte slices are allocated
in one contiguous block to reduce allocations.

The new encoder and decoder are about 5x faster with 10x fewer allocs.

+ benchmark

internal/compile/compile.go

@@ -37,7 +37,7 @@ import (
 const debug = false // TODO(adonovan): use a bitmap of options; and regexp to match files
 
 // Increment this to force recompilation of saved bytecode files.
-const Version = 3
+const Version = 4
 
 type Opcode uint8