fix(nvarc): bound and reap interpreter processes

[ananthsub]

Summary

• cap inductive Python subprocesses per server worker with a positive python_max_concurrency setting
• terminate, escalate to kill, and wait for children on timeout or caller cancellation before returning
• expose queue and execution time as separate verify spans and document timeout admission semantics
• replace duplicated test implementations with production-path tests covering success and every failure lifecycle

Correctness and reliability

The configured execution timeout starts after semaphore admission, so queueing does not consume a submission's runtime budget. Existing None prediction and zero-reward behavior is preserved for execution failures. This change makes no throughput improvement claim; admission control intentionally trades queue latency for bounded resource use.

Validation

• uv run --extra dev pytest resources_servers/nvarc/tests/test_app.py -q — 43 passed
• scoped pre-commit — passed
• real-process stress: 32 infinite-loop submissions at concurrency 8 produced peak 8 children, zero survivors after quiescence, and 4.11 s elapsed
• repeated-failure test checks both descendants and open file descriptors after 32 subprocess failures

Closes #3029

[ananthsub]

fyi @esarafian

[cmunley1]

should this stay inside the try?
If spawning raises OSError (for example, process or file-descriptor exhaustion), verification now returns a 500 instead of the previous None