Table allocation

Many things conspire to make this tricky:

1. Any table field can be omitted
2. Tables encoding should be canonical
3. Tables are encoded as length-prefixed array of offsets to fields

Some possible solutions:

- Users must declare up-front how many fields they will encode. Tables are encoded as a packed array of `(index, offset)` tuples. Accessing a table requires doing a binary search on the index.

- Users must declare up-front the index of the last field they will encode. Tables are encoded as an array of `offsets`, with zero-offsets for missing entries. Accessing a table requires checking the length, so that you know how many fields are present. Fields that would be past the end of the array are considered to be missing.

- Buffers are allocated back to front, so the table can be allocated after the fields are allocated, so you don't need to know how many fields are present up front.