[build] Add a local development kit for contributors

[Gezi-lzq]

Purpose

Closes #3850.

This draft provides a concrete implementation of a contributor-oriented DevKit for running and validating locally built Fluss artifacts. It is opened as a draft to collect feedback on its scope, dependencies, and maintenance boundary.

• No generative AI tools used
• Yes, OpenAI Codex was used to assist design, implementation, testing, and review

Brief change log

• Add a Docker Compose and just based local development environment
• Mount the local build-target without building a Fluss image
• Include Flink in the core environment while keeping Lake Tiering profile-specific
• Support one or three TabletServers and Core, Iceberg, Paimon, Lance, and Hudi profiles
• Add composable commands for startup, logs, host SQL execution, tiering validation, and cleanup
• Keep filesystem plugins isolated from Lake plugin classpaths
• Run Fluss processes under Docker init and expose JDWP and Prometheus metrics endpoints
• Propagate configured Hadoop settings to the Hudi write client so the S3A-backed profile retains its endpoint and credentials

Tests

The DevKit was exercised manually with both single-node and three-node core clusters. All profile Compose configurations rendered successfully, just --list exposed the expected public commands, and host SQL files were executed through just run-sql.

After the final dependency and profile simplification, a clean-volume regression started the core Fluss and Flink services to healthy state and confirmed that the core profile does not submit a Tiering Job. A host SQL file connected through the Flink SQL Client and returned the Fluss database. The Paimon profile then completed table creation, a three-row write, a running Tiering Job, and a Lake query returning three rows with ID sum 6 and payload range alpha to gamma.

Clean core clusters reached healthy state with JDWP reachable for every Fluss process and each Prometheus endpoint returning fluss_* metrics. A dedicated runtime probe confirmed that Docker init runs as PID 1 with the Fluss Java process as its direct child, and normal Compose shutdown completed successfully. Iceberg, Paimon, and Hudi each completed the create-table, write, tier, and lake-query workflow. Lance was validated through table creation, writes, and a running tiering job. Maven validation, Checkstyle, Spotless, and RAT checks also passed.

Full mvn verify has not yet completed in the current local environment, so this PR remains a draft.

API and Format

No public API or storage format changes.

Documentation

Add devkit/README.md covering setup, profiles, operations, validation, debugging, metrics, and profile extension.