I added two variants, one which is unbalanced (always fusing a single arena into the group) while the other is balanced (always fusing two groups of equal size).
Unfortunately it is difficult to benchmark `upb_Arena_Free()` separately from fuse, since the only way to get a non-zero refcount is to fuse.
Results on my machine:
```
CPU: Intel Skylake Xeon with HyperThreading (48 cores) dL1:32KB dL2:1024KB dL3:38MB
Benchmark Time(ns) CPU(ns) Iterations
-------------------------------------------------------------------
BM_ArenaFuseUnbalanced/2 52.2 52.2 1000000 38.300M items/s
BM_ArenaFuseUnbalanced/8 416 416 142315 19.243M items/s
BM_ArenaFuseUnbalanced/64 4034 4033 14306 15.868M items/s
BM_ArenaFuseUnbalanced/128 8302 8301 6837 15.420M items/s
BM_ArenaFuseBalanced/2 54.7 54.7 1000000 36.581M items/s
BM_ArenaFuseBalanced/8 425 425 100000 18.845M items/s
BM_ArenaFuseBalanced/64 4029 4029 14632 15.886M items/s
BM_ArenaFuseBalanced/128 8050 8049 7176 15.902M items/s
```
PiperOrigin-RevId: 518292073
Protobuf Team Bot
Joshua Haberman
Protobuf Team
Joshua Haberman