fix(thrift): use standard XOR zigzag decode to fix numeric message offset

[EdamAme-x]

Summary

• Fixed TMoreCompactProtocol.decodeZigZag() which used a multiplication-based formula (n >> 1) * (sign) instead of the standard XOR formula (n >> 1) ^ -(n & 1). The old formula was off by +1 for every negative value (odd encoded integers).
• Fixed the type-16 string-ID delta decoder which used the same broken formula. Since lastStringId accumulates across the entire parse, each negative delta error compounded — causing numeric-only message text (e.g. "100") to arrive as "101", "102", or "103".
• The offset grew with sending speed because more operations in a single sync batch meant more type-16 string fields with negative deltas, accumulating a larger error.
• Added regression tests for decodeZigZag covering positive values, negative values, and round-trip with standard zigzag encoding.

Closes #200

Test plan

• deno test packages/linejs/base/thrift/readwrite/tmc.test.ts passes (3 tests)
• deno check passes
• deno fmt passes
• Manual verification: send numeric-only messages ("100", "200") at various speeds and confirm message.text matches exactly

🤖 Generated with Claude Code

Co-authored-by: pchrphhee pchrphhee@users.noreply.github.com

[EdamAme-x]

thanks you @pchrphhee