[bitnet] Full #346 conformance: BitNetWeightLoader/RuntimeWeights naming + llm-runtime/kbitnet facade

[michalharakal]

Motivation

#346 names the BitNet module as "the first greenfield instantiation of the template", but the folded-in module (PR #353) deviates from the convention on several rows. Since BitNet is meant to be the module new families copy, it should exemplify the template fully.

Current deviations (Apertus is the reference shape per #346):

#346 row	BitNet today
<F>NetworkDef.kt	✅ BitNetNetworkDef.kt
<F>NetworkLoader.kt	✅ BitNetNetworkLoader.kt
<F>WeightLoader.kt	❌ BitNetPackedGgufLoader.kt; generic path reuses llama's shared DecoderGgufWeightLoader
<F>RuntimeWeights.kt + <F>TensorNames	❌ reuses DecoderGgufWeights + LlamaModelMetadata; names live in BitNetGGUFNameResolver.kt
<F>ConfigParser	✅-by-design — shared decoderMetadataFromGguf is the #346 outcome, not drift
Runtime facade <F>Ingestion in llm-runtime/k<f>	❌ no llm-runtime/kbitnet exists

Tasks

• Rename BitNetPackedGgufLoader → BitNetWeightLoader (+ API dump)
• BitNetRuntimeWeights.kt: family weight containers + BitNetTensorNames — may delegate internally to the shared decoder containers, but the public surface follows the template
• New module llm-runtime/kbitnet with a BitNetIngestion facade mirroring llm-runtime/kllama
• Anything deliberately kept shared documented as a justified deviation in a PR against Common structure and naming convention for model family modules #346, per that issue's own rule

Verification

• Module file listing matches the Common structure and naming convention for model family modules #346 table one-to-one (or carries a documented deviation).
• Existing bitnet tests green after the renames; a kbitnet ingestion round-trip test.

Refs: #346, #336, #335.