perf: bulk calls to Otter

[dunglas]

I tried using the bulk feature of Otter to improve performance when dealing with uncommon headers. According to the benchmark I introduced in #2052, this slightly improves performance but at the cost of more allocations. I'm not sure if it is worth it.

goarch: arm64
pkg: github.com/dunglas/frankenphp
cpu: Apple M1 Pro
                   │  bench.old  │             bench.new              │
                   │   sec/op    │   sec/op     vs base               │
UncommonHeaders-10   131.4µ ± 2%   127.0µ ± 2%  -3.39% (p=0.004 n=10)

                   │   bench.old   │           bench.new            │
                   │     B/op      │     B/op      vs base          │
UncommonHeaders-10   19.31Ki ± 41%   19.40Ki ± 2%  ~ (p=0.796 n=10)

                   │ bench.old  │             bench.new              │
                   │ allocs/op  │ allocs/op   vs base                │
UncommonHeaders-10   19.00 ± 0%   22.00 ± 0%  +15.79% (p=0.000 n=10)