[MP] Various performance optimizations

[Faless]

As pointed out in #81593 , watched/on change variables are quite heavy on the CPU side, mostly due to the way network IDs are exchanged and stored for non-spawned node.

This PR introduce a series of optimizations to reduce the computational cost of the replication system:

• Optimize the data structures used by the network IDs cache (using ObjectIDs locally, while still transferring relative paths).
• Optimize internal authority checks to avoid going through the SceneTree (searching the appropriate MultiplayerAPI as we already know the correct one)
• Avoid unnecessary ref/unrefs between the various internal components (Replication/Cache/RPC)
• Optimize internal SceneReplicationConfig access (using direct pointers instead of refs)

Closes #81593 , might need some more testing.

[akien-mga]

CC @DennisManaa

[DennisManaa]

Many thanks for addressing this issue :) This PR looks great 🎉

Unfortunately, I won't be able to look into this topic until next Tuesday/Wednesday because I'm currently on vacation. But it will be the first thing I'll do when I'm home again :)

[DennisManaa]

@Faless, I just wanted to let you know that I can confirm the performance benefits of your changes. Additionally, I didn't notice any further problems while testing them with our game.

I measured the cycles for the last commit in this PR and for the last commit before the changes in this PR (commit 0ca8542). It seems that you reduced the needed cycles for _send_delta to approximately 31% and the needed cycles for _send_sync to approximately 41% in our specific project. Many thanks for your great work! ❤️
(I think that others should probably experience similar performance-improvements, when using a lot of MultiplayerSynchronizers in their projects) :)

Here are the results of my measurements:

Function	before	after	"after" divided by "before"
Main::iteration	1.714E+11 cycles	1.262E+11 cycles	≈ 0.74
SceneTree::process	1.385E+11 cycles	0.9372E+11 cycles	≈ 0.67
SceneMultiplayer::poll	0.8381E+11 cycles	0.3825E+11 cycles	≈ 0.46
SceneReplicationInterface::on_network_process	0.8064E+11 cycles	0.3475E+11 cycles	≈ 0.43
SceneReplicationInterface::_send_delta	0.4717E+11 cycles	0.1448E+11 cycles	≈ 0.31
SceneReplicationInterface::_send_sync	0.2401E+11 cycles	0.09872E+11 cycles	≈ 0.41

Before

After

[akien-mga]

Just did a code style review.

This was successfully tested, and the implementation changes seem sensible, so I think this can be merged.

[akien-mga]

Thanks!