perf: Replace pushback with rewinding

[kitten]

Summary

Instead of pushing back actual data and storing it, we can instead use the fact that we're only replaying bytes that have just been returned. This allows us to rewind() a few bytes, instead of copying them into a buffer. This is possible due to block.copyWithin. For cases where we only returned input bytes directly, this is even simpler, since we can store an offset, avoid subarray, and subtract from the offset.

Set of changes

• Always allow partial returns on read()
• Switch input on reader.ts to be offset based and defer subarray calls
• Replace pushback() logic with rewind()
• [unrelated] Replace copyUint8Array with slice
• [unrelated] Share intermediary buffer copy for readUntilBoundary on reader
• [unrelated] Pass in skip table externally as arguments
• [unrelated] Replace expectTrailer with direct read() call
• Add missing copyless return optimization to buffered input path, which is now possible