Implement device top-k logits transfer reduction proof

[EffortlessSteven]

Follow-up from #6010.

#6010 records the current dense Qwen CUDA logits-transfer blocker in receipts and keeps transfer-reduction, speedup, full-residency, broad server, Qwen3, and BitNet QK256 claims false.

Goal:
Implement a real device-side top-k logits transfer path for dense Qwen CUDA so the receipt can show reduced device-to-host logits bytes without weakening answer evidence.

Acceptance:

• Actual D2H logits bytes are lower than full logits for the same generated-token count.
• Receipt preserves requested top-k, selected-token, generated-token, vector-length, and byte-accounting evidence.
• logits_transfer_reduction.claimed=true remains invalid unless measured reduction is present and accounting floors are coherent.
• speedup_claim, full_residency_claim, and server readiness stay false unless their separate exact-profile gates are satisfied.
• No Qwen3, BitNet packed I2_S/QK256, broad CUDA, or product speedup claim is promoted by this optimization alone.
• Support bundle / receipt explanation can point to the blocker or the measured reduction without requiring users to inspect raw logs.

Proof should include focused receipt validation plus the exact dense Qwen CUDA command/profile used to generate the transfer receipt.