fix(context): avoid Context.Clone deadlock under a concurrent writer

[Shinku-Chen]

Context.Clone held c.lock.RLock and then called c.ForEach, which
acquires c.lock.RLock a second time on the same goroutine. sync.RWMutex
intentionally starves additional readers once a writer is queued (to
prevent writer starvation), so the second RLock could block forever,
deadlocking Clone whenever any other goroutine had a Put queued
between the two RLock calls.

Iterate c.contextMap directly under the single RLock instead, copying
each entry into the freshly constructed (uncontended) new context. No
behavior change for callers; the lock acquisition count drops from
two to one.

Add TestContextCloneConcurrentWriter, which fires 500 Clone calls and
500 Put calls in parallel under a 5s deadline. The test reliably
times out on the previous implementation and completes in tens of
milliseconds with the fix.