Steer prompts into the running turn instead of queueing them

[OSadovy]

Implements the steering proposal from #773.

When the agent advertises steering and a turn is running, a prompt sent from the
shell is injected into that turn instead of being queued until it ends. Agents
that don't advertise it queue exactly as before.

The method is _session/steering — an ACP extension rather than spec (hence the
leading underscore), advertised in the initialize response's top-level _meta
as steering.supported, a sibling of agentCapabilities. Verified against
@agentclientprotocol/claude-agent-acp 0.66.0 and @agentclientprotocol/codex-acp
1.2.0. The standards-track equivalent, session/inject
(agentclientprotocol/agent-client-protocol#1261), is unmerged and targets ACP v2,
so the extension is what's usable today.

Queueing stays the fallback and nothing is lost. A steer is attempted only
when the agent advertises it, agent-shell-steer-when-busy is on, and a turn is
running that isn't blocked on a permission answer. Every other outcome — the turn
had ended, the agent refused, the request failed — either submits the prompt
normally or puts it back on the queue. A shell waiting on a permission answer
keeps queueing deliberately: what an agent does with a message injected while a
tool sits on that question isn't defined by any implementation I could find.

Entry points:

• Sending from a busy shell steers automatically.
• C-u M-x agent-shell-prompt-queue queues that one prompt instead.
• M-x agent-shell-steer steers explicitly and errors when it can't.
• agent-shell-steer-when-busy (the one new defcustom, since [feature] steer prompts into the running turn #773 asked for
  opt-out) turns the automatic behaviour off entirely.

Neither adapter echoes a steered prompt back while the turn runs, so it's rendered
client-side, reusing the field/face shape already used for replayed user messages
and closed with shell-maker's end-of-prompt marker — without that, chat mode reads
it as the live prompt and the prompt bar hides it. Also adds a prompt-steered
session event.

On automating cancel-and-resend instead (your comment in the issue): I don't
think they're equivalent. Interrupting drops the work in flight and leaves a
cancellation in the transcript, while a steer keeps the turn and what it has
already done. They compose, though — agent-shell-steer could fall back to
interrupt-and-continue for agents that don't advertise steering, which is what
@liaowang11 suggested. Happy to do that as a follow-up if you'd like it.

Behaviour differs per agent, worth knowing before trying it: Claude interrupts
what it's generating and answers in a couple of seconds; Codex accepts the steer
but finishes the message in flight first.

Testing: five tests covering the request shape, outcome mapping, the decision to
steer, the four-way fallback (including that a failed steer keeps the prompt), and
the rendered prompt. Full suite passes at 568.

Checklist

• I agree to communicate (PR description and comments) with the author myself (not AI-generated).
• I've reviewed all code in PR myself and will vouch for its quality.
• I've read and followed the Contributing guidelines.
• I've filed a feature request/discussion for a new feature.
• I'm making visual changes, so I'm including screenshots so you can view and discuss.
• I've added tests where applicable.
• I've updated documentation where necessary.
• I've run M-x checkdoc and M-x byte-compile-file.