[alvarobartt]

Description

This PR improves CUDA serving for Qwen3.5 decision models with fused operations, packed batching and memory-aware token admission. Model heads reuse the shared architecture code, with CPU and Metal implementations kept separate.

Cloudflare/clef-flash BF16 (17f0b0a), one NVIDIA H200; sys1 16965e3, SGLang 0.5.23. Two runs per engine: 748 requests per curve point; 17,952 requests per sustained run.

Concurrency	Engine	Mean request (ms)	p95 (ms)	Requests/s
1	sys1	34.64	89.94	28.79
1	SGLang	55.75	98.10	17.89
4	sys1	113.76	289.62	35.11
4	SGLang	137.58	295.19	29.05
16	sys1	430.40	692.41	37.07
16	SGLang	445.08	722.31	35.86
32	sys1	834.51	1,147.65	38.08
32	SGLang	834.58	1,155.36	37.81
64	sys1	1,615.77	2,104.03	38.66
64	SGLang	1,627.23	2,195.67	38.11
128	sys1	3,085.89	4,901.79	38.83
128	SGLang	3,185.50	4,412.97	38.18
128 sustained	sys1	3,281.88	5,296.82	38.86
128 sustained	SGLang	3,335.62	4,163.64	38.27
512 sustained	sys1	12,964.54	15,511.16	38.87
512 sustained	SGLang	13,188.64	14,420.82	38.30