feat: repgraph-driven per-client deltas

[zachsmith1]

Summary

• Use repgraph to build per-client deltas (no per-client filtering/cloning).
• Reuse per-client encoder state + buffers (delta/full/compact) to cut allocations.
• Add BevySchema::build_delta_update for repgraph world adapter.

Benchmarks (baseline -> optimized)

Scenario A (Relevancy, 1024 ents, 128 clients, 10% dirty)

• bytes_avg: 254 -> 254
• encode_us_avg: 433 -> 411 (~5% better)
• encode_us_p95: 554 -> 498 (~10% better)

Scenario B (Dense, 512 ents, 1 client, 100% dirty)

• bytes_avg: 5515 -> 5515
• encode_us_avg: 1440 -> 1806 (worse; graph overhead dominates single-client)

Scenario C (Loss, 512 ents, 64 clients, 10% dirty, drop 3%, reorder 5%)

• bytes_avg: 129 -> 129
• encode_us_avg: 334 -> 202 (~40% better)
• encode_us_p95: 452 -> 243 (~46% better)

Notes

• This shifts per-client work to "encode-only" and reuses buffers.
• Dense single-client case shows repgraph overhead; multi-client cases improve.

Tests

• cargo clippy --all-targets --all-features