feat: implement window mechanism to eliminate memmove operations

[yc199911]

Problem: pktbuf_is_complete() and pktbuf_pop_packet() use expensive memmove()
operations for packet alignment, causing O(n) performance overhead when
processing packets with garbage prefix data.

Changes:

• Add 'start' field to pktbuf_t for window-based buffer management
• Refactor pktbuf_is_complete() to use window offset instead of memmove
• Refactor pktbuf_pop_packet() to use window offset instead of memmove
• Change packet alignment from O(n) to O(1) time complexity

Future Considerations

This change lays groundwork for potential ring buffer implementation with readv() zero-copy I/O. Would appreciate maintainer feedback on this direction.

[RinHizakura]

Sure, using memmove() is not the best solution. But does this really work? When a packet is consumed, how can you recycle the buffer space? It seems like you just keep increasing the buffer to accept more incoming data.

If that's the case, the statement "Change packet alignment from O(n) to O(1) time complexity" actually means nothing, because it involves more memory allocation, which means paying extra cost in the global perspective for such "local optimization.

The ideal solution is to design a ring buffer for the packet buffer.