perf: rewrite RlpNode internals with manual u8 length + MaybeUninit buffer

[gakonst]

Summary

Rewrite RlpNode internals to use a hand-rolled u8 length + [MaybeUninit<u8>; 33] buffer instead of ArrayVec<u8, 33>.

Motivation

ArrayVec uses a u32 length field (4 bytes + padding), wasting space for a type that can hold at most 33 bytes. Switching to u8 length shrinks RlpNode from 40 to 34 bytes and makes it Copy.

Changes

• Replaced ArrayVec<u8, 33> with u8 len + [MaybeUninit<u8>; 33]
• RlpNode is now Copy — clone is a plain memcpy
• from_raw, word_rlp, as_slice, as_mut_slice are now const
• Removed arrayvec direct dependency (kept as transitive via alloy-rlp)
• Added serde bridge via From impls

Testing

All 34 existing tests pass.

Prompted by: danipopes

[codspeed-hq]

Merging this PR will improve performance by 14.35%

⚡ 1 improved benchmark
✅ 3 untouched benchmarks

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
⚡	trie[64]	464.7 ns	406.4 ns	+14.35%

---

_{Comparing danipopes/rlpnode-internals (b0bf712) with main (6e42e13)}