Fix nonblocking response write handling and backpressure

[AlexJuca]

Context

FKVS uses nonblocking sockets and buffers outgoing responses in client_t.wbuf, but the event loops currently watch read readiness only. If send() returns EAGAIN or only writes part of a response, bytes can remain in wbuf without any write-readiness registration to flush them later.

Relevant code:

• src/commands/common/command_registry.c (wbuf_append, wbuf_flush, direct send() fallback)
• src/networking/networking.c (try_process_frames flushes once after reads)
• src/io/event_dispatcher_epoll.c
• src/io/event_dispatcher_kqueue.c
• src/io/event_dispatcher_io_uring.c

Scope

• Track pending response bytes per client.
• Register write readiness (EPOLLOUT/EVFILT_WRITE or io_uring send operations) when wbuf_used > 0.
• Disable write readiness once the buffer is fully flushed.
• Remove or make safe the direct send() fallback in wbuf_append; all writes must handle partial sends.
• Decide how to handle clients that stay backpressured too long.

Acceptance criteria

• Large responses and pipelined command bursts do not stall under receiver backpressure.
• Partial writes are retried without response corruption or reordering.
• Tests cover forced partial writes / EAGAIN behavior.
• No unchecked send() path remains for framed responses.