Fix agent sandbox creation and process bridge failures

[derekdeming]

Summary

• serialize agent-scoped sandbox initialization so concurrent conversations cannot create multiple supposedly shared sandboxes
• reject unsuccessful process-bridge responses for output polling, stdin writes, and stdin closure instead of spinning or silently accepting the failure
• add one regression test covering process-bridge error termination

Validation

• cargo test -p exoharness --features basic-backend -p executor
• cargo clippy --workspace --all-targets -- -D warnings
• oxlint --deny-warnings .
• tsgo --noEmit -p tsconfig.json
• vitest run

[derekdeming]

Personally noticed two issues while playing around with the harness. Exo could accidentally create 2 sandboxes for the same agent when conversations started simultaneously. Then it could also ignore a failed response from aws agentcore or firecracker and poll forever. The PR ensures each agent gets only 1 shared sandbox and stops failed remote commands immediately