[codex] experimental python 3.14t feature server

[franciscojavierarceo]

What this changes

This PR adds an experimental Python 3.14 free-threaded feature server container and a benchmark harness for comparing it against the current image.

It also trims eager pandas imports from the read-serving startup path so the experimental server can boot without immediately falling back to GIL-protected native modules.

Highlights

• adds a python314t experimental feature-server Dockerfile and requirements set
• adds a benchmark script for stable vs experimental feature-server comparisons
• updates build targets and Docker context handling for the experimental image
• lazy-loads pandas and related offline-only paths across the serving startup path
• updates the feature-server and development docs to describe the experimental image, benchmarking workflow, and current SQLite-oriented scope

Why

We want a safe, clearly labeled place to evaluate whether free-threaded Python can improve feature-server performance without changing the default production image.

The import cleanup is part of making that experiment meaningful: the server startup path should avoid importing pandas unless an endpoint or workflow actually needs it.

Validation

• python3 -m py_compile on the touched Python modules
• ruff check on the benchmark script and lazy-import follow-up fixes
• built feastdev/feature-server:local-b15df7901-python314t-experimental
• verified import feast.feature_server in the experimental image leaves pandas unloaded
• started the experimental server against the SQLite sample repo and confirmed the health check succeeds

Notes

• this remains an experimental path and is documented as such
• the broader FeatureStore offline path still imports pandas through offline-store modules, so this PR focuses on the feature-server startup path rather than the entire runtime graph