feat(qwen4_exp): support channel-wise fp8 dense weights and per-row PLE loaders

[alvarorsouza-arch]

What

Adds loaders for the huginnfork/primitive-ai checkpoint variant of Qwen3.8-Flash-Next, which stores the dense backbone in channel-wise fp8 (separate per-channel scales) instead of the fused layout, and ships its PLE table in a per-row format.

• weight.py: fuse the channel-wise scale vectors in the expected order at load time
• ple.py / kernel/triton/ple.py: per-row PLE table loader
• model.py: wire the new loaders

Why

The un-quantized backbone of this checkpoint is ~11 GB of BF16. On a 16 GB card there is no room for a meaningful MoE cache with it. The fp8 backbone is 8.2 GB, which frees enough VRAM to enable --moe-prefill-overlap.

This addresses #409 (12 GB cards: ~9.9 GiB of unquantized attention/embedding weights leave no room for MoE cache + KV) for the 16 GB tier, the same way #390 did for the qwen3_5_moe family.

Measured (RTX 4070 Ti SUPER 16 GB, EPYC 7K62, 251 GB RAM)

	before (bf16 backbone)	after (fp8 backbone)
Backbone on GPU	11 GB	8.2 GB
Prefill	424 t/s (no overlap)	~1000 t/s (overlap on)
Decode	~30 t/s	37 t/s
Quality (40 known-answer QA + 30 cloze, temp 0)	39/40	39/40

Identical quality score against the same suite, A/B on the same server.

Test protocol

All numbers measured end-to-end over the OpenAI API. The quality suite is 40 known-answer questions plus 30 completions at temperature 0, A/B against the bf16 checkpoint on the same hardware. Happy to add a loader unit test if maintainers prefer.