fix(ci): per-SHA reset and dispatch race fix in CI poller

[chihsuan]

Context

CI escalation fired on the same poll that dispatched the agent, killing a 43-second run before it could fix anything (RSM-3739, PR #57). The lifetime ci_retry_count also stuck issues in escalated across new commits.

TL;DR

Reset CI dispatch/rerun history per head SHA, protect in-flight dispatches from being escalated, and stop poll_error from clobbering the state-machine status.

Summary

• New head SHA observed → clear dispatched_shas, rerun_attempted_shas, downgrade escalated → watching. Lifetime ci_retry_count is preserved; only green resets it.
• Reorder handle_ci_failure so a SHA already in dispatched_shas short-circuits to :already_handled before the escalation branch, preventing the next poll from killing the in-flight agent.
• poll_error_attrs no longer writes status: "poll_error" — only :error, :consecutive_errors, :next_poll_at. Statuses like escalated/rerun_requested survive a transient GitHub error.
• Add tests for new-SHA reset, dispatched-SHA race protection, and status preservation across poll errors.
• Update docs/configuration.md to describe the new retry semantics.

Alternatives

• Reset lifetime ci_retry_count on every new SHA — rejected: escalation would never fire when an agent rapidly pushes new broken commits.
• Escalate atomically when retry_count + 1 == max_retries (skip the last dispatch instead of killing it) — rejected: drops one dispatch attempt and is harder to reason about than "protect the SHA you just dispatched."
• Track full per-SHA history (counts, results) — rejected as bigger than needed; the SHA list already dedupes dispatches.

Test Plan

• make all
• mix test test/symphony_elixir/ci_poller_test.exs (18 tests passing locally)