Support additional BitNet models (Falcon-E, BitNet 0.7B)

[m96-chan]

Summary

• Fix activation detection: Only arch=bitnet-25 (BitNet 2B-4T) uses relu²; all other architectures default to silu. Removes broken isOfficial heuristic that misclassified Falcon-E (large vocab) and BitNet 0.7B (arch=bitnet).
• Handle F16 untied LM head: Falcon-E models have output.weight as F16, not I2_S ternary. Keep as raw F16 and route through the f32_matmul shader path instead of BitLinear.
• Add ChatML template: Auto-detect <|im_start|>/<|im_end|> tokens in vocab and apply ChatML format. Also adds <|im_end|> as a stop token.
• Fix BITNET_0_7B_CONFIG: Correct hiddenSize (2048→1536), intermediateSize (5632→4096), numAttentionHeads (32→16), numKeyValueHeads (32→16).
• Add model presets: Falcon-E 1B (~400 MB), Falcon-E 3B (~700 MB), BitNet 0.7B (~200 MB) in web-chat demo.

Closes #1

Files Changed

File	Changes
packages/core/src/types.ts	Add lmHeadF16?: boolean to ModelConfig
packages/core/src/model/loader.ts	Fix activation detection; F16 lm_head handling; skip dummy scale for F16 head
packages/core/src/model/config.ts	Fix BITNET_0_7B_CONFIG dimensions
packages/core/src/nn/model.ts	F16 untied lm_head path; parameterize dispatchLMHead weight buffer
packages/core/src/tokenizer/tokenizer.ts	Add ChatML template; imEndTokenId getter
packages/core/src/index.ts	Stop on `<
examples/web-chat/src/main.ts	Add 3 model presets

Test plan

• npm run build passes
• npm run lint passes
• Load BitNet 2B-4T — verify no regression (relu², LLaMA 3 template, tied embeddings)
• Load BitNet 0.7B — verify silu activation, correct config dimensions
• Load Falcon-E 1B — verify F16 lm_head, ChatML template, <|im_end|> stop token
• Load Falcon-E 3B — same as 1B

🤖 Generated with Claude Code