fix: enforce first-persisted terminal state under the task lock (#401)

[chopmob-cloud]

What

Closes #401. Prevents a racing forced terminal transition (background failure or
cancellation) from overwriting a terminal state another writer has already
persisted. Adds an atomic "first persisted terminal state wins" guard inside
A2AServer.ApplyEventAsync, under the per-task lock.

Why

CancelTaskAsync and TryTransitionToFailedAsync both check IsTerminal before
taking the per-task lock. A concurrent writer can persist a terminal state in the
window between that optimistic check and the locked apply, so the forced writer
then overwrites it:

• Background failure can overwrite a Completed or Canceled task with Failed.
• Cancellation can overwrite a Completed or Failed task with Canceled, so a
  task the agent actually completed is reported as canceled.

How

ApplyEventAsync is the single choke point through which every terminal write is
persisted, and it already holds the per-task lock. Re-reading the current state
under that lock and dropping a status update that would move an already-terminal
task to a different state closes the race for every forced writer at once,
including any added later, rather than patching each call site:

if (currentTask is not null
    && currentTask.Status.State.IsTerminal()
    && response.StatusUpdate is { } racingStatus
    && racingStatus.Status.State != currentTask.Status.State)
{
    return;
}

The existing pre-lock checks are kept as fast-path rejections for the common
(non-racing) case. The guard is scoped to status-update events, which are the
vector used by both forced terminal writers.

Tests

Two deterministic regression tests in tests/A2A.UnitTests/GitHubIssues/Issue401.cs,
one per forced writer. Each drives the race window between the pre-lock check and
the locked apply (the cancel test gates the cancel handler; the background-failure
test arms a task store that injects a terminal state on the drain's under-lock read
and signals when to await the drain safely). Both fail without the guard
(Actual: Canceled and Actual: Failed) and pass with it.

Validation

Clean container, base main, both target frameworks:

• Build clean on net8.0 and net10.0 (0 warnings, 0 errors).
• Full A2A.UnitTests suite: 422 passed / 0 failed on both frameworks (420 pre-existing
  plus the 2 new tests). No regressions.
• Both new tests pass on both frameworks, zero flakiness across repeated and isolated
  runs. With the guard reverted, both fail with the expected actuals, confirming they
  exercise the fix.

Closes #401.