Buffer rework - #622

The old buffer implementation was written in the 90ies when tinyproxy
primarily forwarded plain HTTP requests. It used a linked list of
bufline_s structures optimized for line-based text processing of
HTTP request and response lines.

A typical tinyproxy instance in 2026 serves 99% of requests via the
CONNECT method to HTTPS endpoints. The old design is ill-suited for
this workload: it performs dynamic allocation for each chunk of data
read from the socket, and removes complete lines from the linked list
as they are sent. This design assumes line-oriented text processing
rather than transparent byte copying between sockets.

Replace the linked-list implementation with a contiguous buffer.
The buffer struct is allocated as a single block with the data area
immediately following the struct in memory. This eliminates per-chunk
allocation overhead and simplifies compaction.

Remove add_to_buffer() from the public API. The old line-oriented
API is replaced by direct read_buffer()/write_buffer() calls
through the poll-driven relay_connection() loop.

Add buffer_space() to buffer.h. This function returns the number
of bytes available in the buffer for writing.

Simplify error handling in write_buffer(). Remove the explicit
ENOBUFS/ENOMEM cases. The default error path now handles all
unexpected errors uniformly.

Update relay_connection() in reqs.c to use buffer_space()
instead of buffer_size() when deciding whether to enable
MYPOLL_READ events.

Remove unused buffer.h includes from html-error.c and main.c.
Remove the HTTP_LINE_LENGTH constant from reqs.c.