#2352 serve: preserve explicit false against configured reasoning defaults

Description

When a config or shared setting supplies a reasoning default, `Service.with_shared()` replaces an explicit `reasoning_effort: false` because its absence check uses truthiness. On 0.1.44.1 with config LOW, the existing HTTP/render fixture produces LOW instead of thinking off for that request. The release already documents `false` as off.

Treat only missing, null or empty effort as absent. Add the explicit-false case to the existing request-over-config HTTP regression. Requests that omit effort still use the configured/shared default; ordinary string levels and template kwargs keep their existing precedence. No kernel, sampling, model policy or default change.

Validation: the added assertion fails on exact tag `296b59d2f73c8f04a805df29c2fa97655adace78` with `low != none`, then all 11 `serve.test_effort_paths` methods pass after the one-line fix. Python/Jinja fixture only; no GPU inference. Neither touched Python file builds a CUDA/HIP/SYCL backend.