Clean up a few issues with ARM-optimized varint decoding.

1) Move the check for invalid varint outside handling a >= 32 bit varint. This gives a modest speedup (+0.5%) on the [fleetbench protogen benchmark](https://github.com/google/fleetbench/tree/main/fleetbench/proto) by working around an LLVM issue which inserted a redundant branch.
2) Move ValueBarrier() to parse_context.h so we can delete ForceToRegister(), which does the same thing as the single value ValueBarrier() implementation.

PiperOrigin-RevId: 504883550