Stream replay loading using seperate Decompression and Seserialize tasks

[tom-leys]

• Improves memory locality by immediately processing replay data that just got loaded into checkpoints (no time to flush from caches & page to disc)
• Reads data in parallel to checkpoint processing so 2x faster by better using multiple CPU cores
• Improves https://github.com/space-wizards/space-station-14/issues/28052 by ensuring low memory systems don't have to do a second pass over replay data that has likely been paged out

How it works

• A decompression task opens and decompresses replay files, handing those MemoryStreams on using a channel
• A deserialization task processes each replay file, creating states and passing those on using a channel
• Checkpoint processing remains in the main thread, reading states as they are deseralized and processing these

On my system, the deserialization task remains a bottleneck (after other Replay Prs merged). We could deserialize with two tasks instead of one to further improve performance but I didn't want to over-complicate this PR.

There's also reuse of decompressed memory stream memory.

• Respects the Job system, using wrapped async waits if there is a Loading Job