Use stack spans for serializer numbers

[JKamsker]

Summary

• switch numeric deserialization in BufferReader to use stack-allocated spans with BinaryPrimitives and keep pooled buffers only for large reads
• mirror the stack span write path in BufferWriter so small numeric writes avoid the array pool while still handling multi-segment copies
• add a regression test that writes and reads numbers across multiple slices to exercise the new stack paths

Testing

• dotnet test LiteDB.Tests -f net8.0

---

https://chatgpt.com/codex/tasks/task_e_68cfd6f1ea80832a91c26680b1edb837