headerKeyCache: use otter v2 (#2040)
Benchmarks show that version 1, while extremely fast with hot keys,
becomes several times slower than no‑cache under frequent misses.
Version 2 delivers consistently better performance across all scenarios,
with no allocations and stable latency.
```
BenchmarkGetUnCommonHeaderNoCacheSequential-12 7545640 169.4 ns/op 72 B/op 4 allocs/op
BenchmarkGetUnCommonHeaderV2Sequential-12 14471982 85.98 ns/op 0 B/op 0 allocs/op
BenchmarkGetUnCommonHeaderV1Sequential-12 19748048 59.63 ns/op 0 B/op 0 allocs/op
BenchmarkGetUnCommonHeaderNoCacheParallelOneKey-12 24352088 44.47 ns/op 72 B/op 4 allocs/op
BenchmarkGetUnCommonHeaderV2ParallelOneKey-12 91024160 11.76 ns/op 0 B/op 0 allocs/op
BenchmarkGetUnCommonHeaderV1ParallelOneKey-12 192048842 6.186 ns/op 0 B/op 0 allocs/op
BenchmarkGetUnCommonHeaderNoCacheParallelRandomMaximumSize-12 26261611 43.07 ns/op 62 B/op 3 allocs/op
BenchmarkGetUnCommonHeaderV2ParallelRandomMaximumSize-12 100000000 14.43 ns/op 0 B/op 0 allocs/op
BenchmarkGetUnCommonHeaderV1ParallelRandomMaximumSize-12 137813384 8.965 ns/op 0 B/op 0 allocs/op
BenchmarkGetUnCommonHeaderNoCacheParallelRandomLenKeys-12 24224664 46.57 ns/op 71 B/op 3 allocs/op
BenchmarkGetUnCommonHeaderV2ParallelRandomLenKeys-12 69002575 17.42 ns/op 0 B/op 0 allocs/op
BenchmarkGetUnCommonHeaderV1ParallelRandomLenKeys-12 8498404 253.1 ns/op 42 B/op 1 allocs/op
```
---------
Co-authored-by: Kévin Dunglas <kevin@dunglas.fr>
