Add LFM2.5 2.6B support and speed up decode

[CerebralCoding]

Summary

• add strict runtime detection and native GGUF inference for the dense LFM2.5-2.6B model alongside LFM2.5-8B-A1B
• derive layer schedules, tensor validation, cache sizing, FFN dimensions, and RoPE configuration from the detected model
• keep decode workers hot between adjacent projections and reserve scheduler headroom for memory-bound matvecs
• bump the project version to 0.2.0 and update the changelog, README, and architecture documentation

Performance

Using the official LFM2.5-2.6B Q4_K_M model and the 256-token benchmark:

• decode: 68.33 → 90.33 tok/s (+32.2%)
• prefill: 201.0 → 199.8 tok/s

Validation

• ReleaseFast unit tests
• ReleaseFast build
• end-to-end tests with LFM2.5-2.6B and LFM2.5-8B-A1B
• LFM2.5-2.6B tool-calling integration
• benchmark continuation correctness check