perf: eliminate frame write fragmentation and optimize buffer management

[countradooku]

• Replace write_vectored with single-buffer writes to avoid fragmentation
• Optimize buffer allocation patterns for small/large frames
• Improve masking performance with 64-bit operations and better alignment
• Add fast-mode constructor with performance-optimized defaults
• Reduce unnecessary buffering in read path for large payloads

Addresses feedback about performance parity with simpler WebSocket implementations. Eliminates the main performance bottlenecks: header/payload fragmentation, excessive buffering, and suboptimal memory access patterns.