Arena allocator

[cBournhonesque]

Context

There are a lot of situations where we suddenly need to allocate a bunch of small objects that we will then discard:

• when buffering all the updates that will go into the replication messages
• when we do priority filtering
• when build the final packets
• when we read the received packets into data structures

It might be beneficial to use an fast arena allocator that allocated a bunch of memory. We use this allocator in the hot path, and then regularly we just cheaply drop the chunk of memory at once.

Solution

Introduces a ArenaManager that wraps a SyncBlinkAlloc from the blink-alloc crate, which provides a multi-thread safe arena allocator.

Use this arena everytime we send to allocate the pending_actions and pending_updates that we need to build the final ReplicationMessage.

Benchmarks

send_insert_float/1_client

num_entities	main	branch
1000	1.30ms	1.21ms
10000	13.10ms	13.23ms

Again it doesn't seem to be bringing much improvement, maybe the allocations in replication_send are not the problem, rather it's the ones in the packet building steps? Double-check.

Anyways it's nice to have a working implementation