`ctl run-benchmark --local` OOM on local infer

our inference process has the tendency to OOM given it loads full git repos into memory. there are various ways we could mitigate this (e.g. with file-backed storage at the cost of speed) but the true fix would be to move inference to a separate process and detect/recover from OOM.

one additional thing to note is that we might want to refactor the inference server (and others) to accept a port flag instead of hard-coding the Cloud Run default. for ctl, we can use the port picker logic from internal/firestoretest/emulator.go to pass in an available port.

And thanks @giacomobenedetti for reporting this issue!

Start with the ctl run-benchmark --local entry point and inspect the inference server; use internal/firestoretest/emulator.go as the reference for selecting an available port. Done should address local inference OOM through process isolation and recovery, while allowing the inference server to accept a chosen port instead of the Cloud Run default.

Written by the indexing model from the issue text.