Expose server proxy and TLS configuration, and fix regressions uncovered by not-slow tests

[applesauce49]

Summary

This PR delivers two things:

1. Adds the server networking configuration feature so operators can run oMLX behind enterprise proxy and TLS-intercepted environments more reliably.
2. Fixes runtime and test regressions found in the not-slow suite, including a real thinking-budget state bug and a boundary snapshot cleanup race.

Feature Added

• Exposed server network controls through the serve configuration flow:
  • HTTP proxy
  • HTTPS proxy
  • No-proxy bypass list
  • Custom CA bundle
• Ensured these settings are applied consistently at startup so downstream requests inherit the expected network/TLS behavior.

Bug Fixes

• Fixed thinking budget token progression logic so forced close-think and natural end detection advance correctly across decode steps.
• Fixed a boundary snapshot async-writer cleanup race where canceled requests could still leave behind request directories or late-written files.
• Hardened cancellation handling in the snapshot writer by re-checking cancellation around file materialization and rename paths.

Test/Compatibility Fixes

To align tests with current intended behavior and interfaces:

• Updated VLM discovery regression fixture to include vision metadata where required.
• Updated CLI scheduler option expectations to current serve flags.
• Updated server health-check tests to patch session-based requests flow.
• Updated VLM tool-parser availability test to reflect the current dual-backend parser fallback path.

Validation

• Re-ran the previously failing set from the not-slow run.
• Result: all 10 previously failing tests now pass.

Impact

• Improves deployability in proxied/TLS-managed environments.
• Resolves runtime correctness issues in thinking-budget control and boundary snapshot cleanup.
• Restores test reliability by matching tests to current production behavior.