fix: let user params override engine defaults in API VLM engine (#3116)

* fix: let user params override engine defaults in API VLM engine

User/model spec params now take precedence over engine defaults
(temperature, max_tokens). This enables Azure OpenAI compatibility
where max_completion_tokens must be used instead of max_tokens, and
temperature=0.0 may be rejected by some deployments.

When max_completion_tokens is set in user params, conflicting
max_tokens is automatically removed.

Fixes #3112

Signed-off-by: majiayu000 <1835304752@qq.com>

* test: add coverage for user param override in API VLM engine

Signed-off-by: majiayu000 <1835304752@qq.com>

* fix: remove mock unit tests per reviewer feedback

Signed-off-by: majiayu000 <1835304752@qq.com>

---------

Signed-off-by: majiayu000 <1835304752@qq.com>