Refactor.

Previously we converted spans from UTF-8 to UTF-16 in AST first, then serialized tokens. This meant that when serializing tokens, we had to convert spans to UTF-16 before comparing offsets of overrides to the token's start offset (because the override offsets are taken from the AST nodes during the AST walk pass).

Instead, convert token spans first, then the AST. This removes span conversion from the critical path. On a CPU with instruction-level parallelism, other work can start without being blocked on waiting for the result of span conversion, and the two can happen in parallel.

This should be a slight perf improvement, though it likely won't show up on CodSpeed benchmarks as CodSpeed's CPU simulation to basic to emulate instruction-level parallelism.

But mainly, the motivation is that this makes the process simpler and easier to reason about.