Refactor replication to use a buffer cache

[UkoeHB]

Problem

When servers build a replication diff, they allocate many times to collect an intermediate representation of the message. This is quite inefficient.

Solution

Use a buffer for the intermediate representation so allocations are greatly reduced. We still have allocations when new entities are introduced.

I saw a 40% speedup in the 'entities send' benchmark, and 6% in the 'entities receive' benchmark. I did not examine packet sizes, although I expect them to be smaller due to some optimizations around entity and replication ids.

This PR depends on #41.

Follow-up

The replication buffer should have a queue of ComponentDatas collected from despawned entities for reuse with new entities. In that case the long-run upper limit on buffer size will be: O(max entities * max entity size).

The buffer currently limits apps to 65k replicable entities and 127 replicable components per entity. It may be possible to configure those values at compile time.

Entities can be serialized as two u32 varints instead of a u64 varint, which should save 1-3 bytes per entity since entities are actually two u32s concatenated.

[codecov-commenter]

Codecov Report

Patch coverage: 96.95% and project coverage change: -0.17% ⚠️

Comparison is base (cbc6825) 95.47% compared to head (0bb7817) 95.31%.

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@            Coverage Diff             @@
##           master      #42      +/-   ##
==========================================
- Coverage   95.47%   95.31%   -0.17%     
==========================================
  Files          10       10              
  Lines         663      725      +62     
==========================================
+ Hits          633      691      +58     
- Misses         30       34       +4     

Files Changed	Coverage Δ
src/replicon_core.rs	96.00% <96.58%> (-0.30%)	⬇️
src/server.rs	95.80% <97.82%> (-0.67%)	⬇️
src/client.rs	84.74% <100.00%> (ø)

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[UkoeHB]

Rebased onto the updated benches. Replicating a DummyComponent(usize) now instead of DummyComponent, so the perf gain from this PR is only 20% for entity creation. However it is 50% faster for updating components (and for some reason the client-side is 45% faster for updating components).

[Shatur]

Superseded by #46.