[codex] Align preloaded model aliases

[hansent]

Summary

• Preserve the requested model alias when startup preloading resolves a registered alias like florence-2-base to florence-pretrains/3.
• Pin preloaded aliased models under the same key that request-time loading uses.
• Add a focused HTTP startup preload test covering the Florence alias behavior.

Root Cause

The request-time LMM path loaded the resolved model ID but passed model_id_alias, causing the model manager and fixed-size cache to key the loaded model by the requested alias. The startup preload path resolved the alias but dropped the original requested ID, so Florence was registered under florence-pretrains/3 while incoming requests looked for florence-2-base.

Validation

• .venv/bin/python -m py_compile inference/core/interfaces/http/http_api.py tests/inference/unit_tests/core/interfaces/http/test_http_api.py
• .venv/bin/python -m pytest tests/inference/unit_tests/core/managers/test_decorators.py tests/inference/unit_tests/core/managers/test_base.py -q

Note: the focused test_http_api.py suite is blocked in this local environment before reaching the new test because importing http_api fails with ImportError: cannot import name 'AutoModel' from 'inference_models'.