Mistral batched decode: LLaMA decode path + parameterized RoPE theta (12x batched, full serving stack)

[mikepapadim]

Mistral batched decode (LLaMA decode path, no new kernels)

Extends the batched-decode engine (PR beehive-lab#129) to Mistral with a tiny generalization — Mistral
is LLaMA-arch (uses LlamaState + LlamaTornadoWeights, all accessors on the base
Configuration), so it runs the existing LLaMA decode path unchanged except:

1. Parameterized RoPE theta in the decode + paged-decode rope kernels (config.ropeTheta();
   Mistral = 1e6, was hardcoded). Prefill ropes untouched; Llama-1B re-verified coherent.
2. Generalized the decode layer graph (LlamaFP16LayersBatchDecodeMMA) to the base
   Configuration and dropped the LlamaConfiguration cast in the engine.

No new kernels. Any LLaMA-family model (RMSNorm + SwiGLU + GQA + RoPE, no QK-norm, no sliding
window) now drops in the same way; the full serving stack (continuous batching, PagedAttention,
prefix caching, on-device sampling) is model-agnostic and applies unchanged.

Verified — Mistral-7B-Instruct-v0.3 FP16, RTX 4090 (CUDA backend)

mode	result
batched B=32	331 tok/s = 12× vs 27 tok/s single-stream; all 32 streams bit-exact, coherent
continuous + paged + prefix (128 requests)	prefix-consistent, coherent; 85% slot utilization
paging	88 / 256 blocks used = ~11.6× less KV than full reservation
prefix caching	82.8% fewer prefill tokens (48-tok shared prompt)

Output sample (greedy): "The capital of France is Paris. It is one of the most famous cities in
the world, known for its art, fashion, cuisine, and historical landmarks such as the Eiffel Tower,
Louvre Museum, and Notre…"

Base branch: feat/static-batched-decode (PR beehive-lab#129). Tracked in roadmap issue beehive-lab#130.