Use [MaybeUninit<u8>] throughout decode and encode paths

[elichai]

• Replace the Vec::set_len before write in decode() and encode_inner() with Vec:spare_capacity_mut(), deferring Vec::set_len until after data is written
• Change all internal decode functions (decode_checked, decode_unchecked, decode_to_slice_inner) to accept &mut [MaybeUninit<u8>] instead of &mut [u8], across all arch backends
• Add Output trait impl for &mut [MaybeUninit<u8>] so SIMD and generic decode paths work with uninit buffers
• decode_to_array now passes its MaybeUninit array directly without casting to &mut [u8] first

[codspeed-hq]

Merging this PR will degrade performance by 19.02%

⚡ 1 improved benchmark
❌ 2 regressed benchmarks
✅ 33 untouched benchmarks

⚠️ Please fix the performance issues or acknowledge them on CodSpeed.

Performance Changes

	Benchmark	BASE	HEAD	Efficiency
❌	bench1_32b	372.5 ns	460 ns	-19.02%
❌	bench2_256b	631.1 ns	747.8 ns	-15.6%
⚡	bench1_32b	580.6 ns	522.2 ns	+11.17%

---

_{Comparing elichai:maybeuninit (63b1bbe) with master (1143588)}

[elichai]

The CI failures seem unrelated (unused portable_simd) and I'm not sure how to interpert the codspeed results, this should not affect performance (by moving the location of .set_len we will probably influence codegen a bit),

I also ran it locally with miri and fuzzed it via cargo +nightly fuzz run fuzz_const_hex --careful for a while on my macOS