Add streaming/tailing variants of log readers #73

Problem

read_error_log and read_access_log are point-in-time reads. For "make a change, hit the page, see what got logged" feedback loops, agents currently have to re-read the whole file each time. A cursor-based tail-with-offset pattern would let agents incrementally consume new log lines as they appear, turning one-shot code generation into an actual debugging loop.

Proposal

• tail_error_log(siteId?, sinceCursor?, maxLines?) → returns new lines since sinceCursor plus a new cursor token
• tail_access_log(siteId?, sinceCursor?, maxLines?) → same shape

Cursor is opaque to the client — implementation can be byte offset, last-line hash, or whatever is robust against log rotation.

Acceptance criteria

• Successive calls with the returned cursor return only new lines
• Handles log rotation gracefully (resets cursor, includes a flag indicating rotation occurred)
• maxLines defaults to a sane value to avoid context blowups
• First call without a cursor returns the most recent N lines plus a cursor for the next call