feat(Add Gemma 4): profile support while keeping Qwen default

[gabsprogrammer]

Summary

• adds Gemma 4 as an additional model-family/profile option without removing the existing Qwen-based defaults
• keeps fresh DreamServer installs on the current Qwen selection path by default
• prepares the installer, tier mapping, and runtime selection logic so Gemma 4 can be enabled explicitly now and supported more cleanly in future backend updates

What Changed

• added MODEL_PROFILE support for qwen, gemma4, and auto
• kept qwen as the default profile for new installs and existing installs that do not opt in
• added Gemma 4 tier mappings across Linux, macOS, and Windows installers
• wired Gemma-aware model metadata into supporting DreamServer services that inspect model names and capabilities
• added newer llama.cpp runtime selection for Gemma 4 paths where needed
• kept auto available as an opt-in profile that can prefer Gemma 4 on stronger tiers while retaining Qwen fallback for the lightest tier

Tier Behavior

Default behavior:

• fresh installs continue to resolve to the current Qwen family by tier
• existing installs continue to behave the same unless MODEL_PROFILE is explicitly changed

Opt-in behavior:

• MODEL_PROFILE=gemma4 enables Gemma 4 tier mappings directly
• MODEL_PROFILE=auto keeps Qwen on the lightest tier and cloud tier, and prefers Gemma 4 on stronger local tiers

Gemma 4 safety tuning:

• tier 3 uses a more conservative context limit for gemma-4-26b-a4b-it
• tier 4 uses a more conservative context limit for gemma-4-31b-it
• SH_COMPACT also uses a safer context limit for the Gemma 4 path
• NV_ULTRA and SH_LARGE keep the longest context configuration

Why

This keeps DreamServer stable for current users while preparing the project for broader model-family support.

Instead of replacing the current Qwen defaults, this PR makes Gemma 4 available as an additional supported path and lays the groundwork for future inference-backend improvements, including newer upstream runtime support and potential TurboQuant-ready backend adoption as those dependencies mature.

Validation

• bash tests/test-tier-map.sh
• tests/bats/bats-core/bin/bats tests/bats-tests/tier-map.bats
• bash -n installers/lib/tier-map.sh
• bash -n installers/macos/lib/tier-map.sh
• bash -n installers/phases/02-detection.sh
• bash -n installers/phases/06-directories.sh
• bash -n installers/phases/08-images.sh
• bash -n installers/macos/install-macos.sh
• bash -n scripts/detect-hardware.sh
• PowerShell tier resolution spot checks on Windows
• Python syntax validation for touched supporting modules

Notes

• this PR does not replace Qwen as the default DreamServer model family
• this PR focuses on DreamServer-side readiness: model profile selection, tier mapping, and compatibility groundwork

[Lightheartdevs]

APPROVE. Clean profile system layered on existing tier architecture. Default is qwen — no change for existing installs. All HuggingFace URLs verified. 9 new BATS tests covering all profile/tier combinations. CI all green. Minor: GGUF_SHA256 empty for Gemma 4 entries (track as follow-up).