Chunk recoder optimization

[cherep58]

// Before optimization
BenchmarkChunkRecoder_min 869701624 ns

// After optimization
BenchmarkChunkRecoder_min 745188236 ns

Boost: ~14%

Re-publishing in English.

[vporoshok]

Thanks for the optimization — the benchmark numbers look nice. However, there are several serious concerns plus a handful of minor ones; see inline comments.

Blockers (must be addressed before merge):

• UB through std::variant in UniversalDecodeIterator::operator* / operator-> — reinterpret_cast over the variant's storage relies on stdlib implementation details. The static_assert below only checks sample_'s offset within each alternative, not the layout of std::variant itself.
• UniversalDecodeIterator::set — same issue but with a write; the method is not even used in this PR.

Important:

• This PR adds but does not use: seek_to, invalidate_sample, set, SeekResult::kUpdateSampleNextAndStop, UniversalDecodeIterator::Type. Better to introduce them in the PR that actually needs them.
• Typo Universale in the public concept name AssignableFromUniversaleDecodeIterator.
• [[likely]] on the kUpdateSample branch in seek is misplaced — the hot path here is ChunkRecoder, which always returns kNext.
• noexcept was dropped from UniversalDecodeIterator::operator== / operator++ — no throwing handler involved there.

Nice to have:

• The LE assumption in bstream.h::write_byte (at minimum, add static_assert(std::endian::native == std::endian::little), or align the style with the existing BareBones::Bit::be(...) usage).
• varint_buffer without {}-initialization in xor.h — needs a contract comment.
• [[likely]] on two consecutive branches in xor.h.
• std::ignore = timestamp_decoder_.decode(); in traits.h.
• The public default constructor of UniversalDecodeIterator creates a "dummy" state.
• The PR description should explain which technique contributes how much — in particular, whether bypassing std::variant is actually required for the +14%.
• Nit: missing trailing newline in traits.h.