Add MiniMax-H3 configurations and config/runner cleanup

[kTorp]

This PR adds MiniMax-H3 configurations and updates the runner to handle new flags. Additionally cleans up some unused configs and code.

Background

MiniMax-H3 is a new and popular model that is updated almost daily in vllm-omni and is of interest to track. The configs are based on the available examples in recipes/MiniMaxAI/MiniMax-H3.md

---

Changes to the online runner due to new MiniMax-H3 configs:

• Added --aspect_ratio, --task_type and --vae_parallel_mode flags used by MiniMax-H3 recipes
• Increased timeouts overall since MiniMax-H3 serving command sometimes timed out during testing
• Updated _parse_json_object string since --diffusion_attention_config is no longer the only json object from configs

Other misc changes to online runner:

• --attention_backend now targets the --diffusion-attention-backend flag instead of the ENV variable
• Simplified server start logging since backend isn't set as ENV variable
• Removed max_sequence_length since it defaults to 512 in Flux.2 (and no need to specify with removal of Flux.1 config)

Cleanups:

• Removed Flux.1 config since it is not actively tracked and also missing upstream sequence parallel support
• Removed offline runner since it is not used by configs and to my knowledge not for any team development either

Tasks

• Add MiniMax-H3 configs
• Update runner
• Misc cleanup of unused config/code

Tests

• Build images
• Test on GFX
• Test on B200
• Test on H100/H200