fix: harden encoder/decoder against UB, data races, and corrupted input

[facontidavide]

Summary

• Fix decodeVarint UB: move shift overflow check before the shift operation to prevent undefined behavior on malformed 11+ byte varints
• Fix decodeVarint underflow: guard against uval==0 (corrupted data) wrapping uint64_t to UINT64_MAX
• Fix worker thread deadlock: detect dead compression worker on subsequent encode() calls instead of hanging forever
• Fix WASM encode performance: compute MaxCompressedSize first and encode directly into caller's buffer when it fits, avoiding unnecessary heap allocation + memcpy
• Clean up worker error semantics: only set worker_failed_ (not compression_done_) in catch block for clearer signal separation
• Optimize encode(BufferView&): use header_.size() directly instead of re-serializing YAML via MaxCompressedSize
• Improve error messages: include field name and type in MaxSerializedFieldSize errors
• Add <cstring> include: make memcpy dependency explicit in encoding_utils.hpp
• Fix FloatLossy test: kResolution was double, causing CTAD to deduce <double> template on float data — changed to float
• Re-enable tests: uncomment test_field_encoders.cpp, test_intrinsics.cpp, test_header.cpp in CMakeLists and remove skip-tests from CI

Test plan

• All 9 unit tests pass (including previously disabled ones)
• Clean build with no warnings
• Pre-commit hooks pass (clang-format, yaml, etc.)
• Verify CI passes on Humble and Jazzy workflows

🤖 Generated with Claude Code