feat(ds4): support qtype 105/106 across heterogeneous GPUs

[davide221]

Summary

• support embedded qtype 105/106 expert weights when DeepSeek 4 is split across two AMD GPUs
• register each GPU-owned expert with its embedded decode table and unregister it before releasing owner buffers
• enable registry-aware mixed-expert prefill on same-runtime GPU tensor parallelism
• scale automatic route balancing with the model's expert top-k, preserving top-4 behavior and supporting the model-default top-6 path

This PR is stacked on #577. Its own diff contains the mixed-quantization support and top-k route scaling; the routing profiles and benchmark framework remain in the base stack.

Validation

Tested with the 0731 mixed-qtype model on an R9700 plus Radeon 8060S, using all six routed experts and the DSpark draft.

• dual-architecture HIP Release build (gfx1151;gfx1201): passed
• heterogeneous planner/storage tests: 10/10 passed
• mixed registry teardown: passed independently on both GPUs
• full DeepSeek 4 GPU unit suite: passed independently on both GPUs
• two warmups plus three measured deterministic 128-token replies
• client decode: 77.184-77.267 tok/s; median 77.221 tok/s
• model-side decode: 68.1-68.2 tok/s
• draft acceptance: 100%
• exact response hash matched on every run

The automatic route split selected 5.0 of 6 expert slots for the faster owner at the qualified 4.4 device-rate ratio.

[cubic-dev-ai]

All reported issues were addressed across 49 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 23 files (changes from recent commits).

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 7 files (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Re-trigger cubic</sub>

[cubic-dev-ai]

All reported issues were addressed across 13 files (changes from recent commits).

<sub>Tip: Review your code locally with the cubic CLI to iterate faster.

Re-trigger cubic</sub>

[davide221]

@cubic-dev-ai review the current head after the branch rebase. Please focus on correctness, maintainability, security, and heterogeneous qtype 105/106 lifecycle safety.

[cubic-dev-ai]

@cubic-dev-ai review the current head after the branch rebase. Please focus on correctness, maintainability, security, and heterogeneous qtype 105/106 lifecycle safety.

@davide221 I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

All reported issues were addressed across 26 files

<sub>Reply with feedback, questions, or to request a fix.

Re-trigger cubic</sub>

[davide221]

@cubic-dev-ai review the current rebased head. Please check correctness, maintainability, security, and the qtype 105/106 heterogeneous routing boundaries.

[cubic-dev-ai]

@cubic-dev-ai review the current rebased head. Please check correctness, maintainability, security, and the qtype 105/106 heterogeneous routing boundaries.

@davide221 I have started the AI code review. It will take a few minutes to complete.

[cubic-dev-ai]

All reported issues were addressed across 25 files

<sub>Requires human review: Auto-approval blocked because this review re-detected 1 unresolved issue already reported by Cubic.

Re-trigger cubic</sub>