fix(common): keep executor alive while contexts exist

[Wondertan]

Summary

• retain the executor shutdown guard in every context created from it
• shut the worker pool down only after the final executor/context owner drops
• cover the reported lifetime with a focused regression test

Root cause

A TLSN SessionHandle owns MPZ's Executor, while a spawned prover owns only a Context. Dropping the handle dropped the executor, which cancelled queued tasks still awaited by that context. async-task::Task then panicked with Task polled after completion.

This is the alpha.6-compatible lifetime fix from upstream MPZ #403, without its unrelated global-pool, concurrency-limit, dependency, or API refactors.

This is orthogonal to #1; it targets #1's base branch and does not alter KOS.

Verification

• cargo test -p mpz-common — 20 passed
• cargo clippy -p mpz-common --all-targets --all-features -- -D warnings
• regression: a context can still run join after its creating Executor is dropped