[refactor] Simplify and unify the TornadoVM layer planner infrastructure

[orionpapadakis]

This PR is a structural refactoring of the tornadovm package — no behavior changes, no new model capabilities. The goal is to eliminate duplication across the FFN layer and layer planner hierarchies and establish a cleaner, more maintainable package structure.

Changes:

FFN layer hierarchy (tornadovm/layers/)

• Introduced AbstractLogitsLayer to centralize shared setup logic for all logits layers (LogitsFP16Layer, LogitsQ8_0Layer, and Granite variants), replacing duplicated task graph and
  scheduler boilerplate
• Generalized AbstractFFNLayers<W, C> with type parameters for Weights and Configuration, giving typed access to weights and config in subclasses
• Unified setupFFNLayerTaskGraphs() across all FFN subclasses; removed redundant fields and methods surfaced during the cleanup
• Added MistralFP16FFNLayers and MistralQ8_0FFNLayers — Mistral-specific FFN layer implementations using MistralConfiguration, required after the generics tightening

Layer planner hierarchy (tornadovm/layerplanner/)

• Centralized createTornadoInferencePlan(), all shared fields (activationLayer, ffnLayers, logitsLayer, scheduler, task graphs), and the GenericLayerPlanner interface implementations
  into QuantizedLayerPlanner, eliminating near-total duplication between FP16LayerPlanner and Q8_0LayerPlanner
• Added MistralFP16LayerPlanner and MistralQ8_0LayerPlanner with correct generic types (LlamaState, MistralConfiguration, LlamaTornadoWeights), fixing a ClassCastException caused by
  Mistral being incorrectly routed to the Llama planners

Package reorganization

• GenericLayerPlanner and QuantizationPlannerFactory moved to layerplanner/ root (were in a base/ subpackage)
• QuantizedLayerPlanner moved to layerplanner/ root
• FP16LayerPlanner co-located with FP16 concrete planners in model/fp16/
• Q8_0LayerPlanner co-located with Q8_0 concrete planners in model/q8_0/

DeepSeek-R1-Distill-Qwen fix

• Introduced DeepSeekR1Qwen model class (extends Qwen2) that correctly overrides getModelType() and shouldAddBeginOfText(), fixing a repetition loop caused by missing BOS token and
  <think> prefix injection
• Fixed Qwen3ChatFormat.getBeginOfText() to fall back to startHeader (<｜begin▁of▁sentence｜>) when no BOS alias is registered — DeepSeek reuses its first role-marker token as BOS
• Fixed Qwen3Tokenizer.encode() to byte-map non-special text parts before BPE encoding, resolving a NoSuchElementException when encoding "\n" after splitting on special tokens like
  <think>