perf: optimize XML hot-part deserialization with streaming and zero-copy

[Nebu1eto]

Summary

Three targeted optimizations eliminating the primary allocation hotspots in the XML read path:

1. Streaming XML deserialization: Replace read_to_string + from_str with from_reader(BufReader) for all XML parts, eliminating multi-MB intermediate String allocations
2. Zero-allocation cell deserialization: Custom serde Visitors for CompactCellRef and CellTypeTag using visit_str(&str) instead of String::deserialize, eliminating 2 heap allocations per cell
3. Zero-copy SST construction: Use std::mem::take to move owned strings from parsed SST directly into Arc<str>, avoiding clone for plain-text entries

Performance Impact (estimated for 100k x 20 workbook)

• ~100MB less peak memory (eliminated worksheet XML String buffer)
• ~2M fewer heap allocations (cell-level deserialization)
• ~10k fewer String clones (SST construction)
• Target: Read Scale 100k >= 8% additional improvement

Changes

• crates/sheetkit-core/src/workbook/io.rs: from_reader(BufReader) with adaptive capacity (8KB-64KB)
• crates/sheetkit-xml/src/worksheet.rs: Custom CompactCellRefVisitor + CellTypeTagVisitor
• crates/sheetkit-core/src/sst.rs: std::mem::take for plain-text SST entries

Test plan

• 6 new tests (from_reader worksheet/SST/large, zero-copy plain/mixed/empty)
• 1,978 Rust tests pass
• clippy clean, fmt clean
• Stacked on Phase 1 (ReadFast mode)

🤖 Generated with Claude Code

[claude]

Code review

No issues found. Checked for bugs and CLAUDE.md compliance.