feat: streaming GGUF → bgz17 indexer (bounded RAM, all model types)

Reads GGUF tensor-by-tensor via seek, projects each weight matrix
to Base17 via golden-step averaging, writes compressed output.
Peak RAM = one tensor + buffers, regardless of model size.

Supports: Attention, FFN, Conv2D, Embedding layer classification.
Conv2D [out_ch, in_ch, kH, kW] reshaped to out_ch vectors of kernel_dim.

14 tests: classification, projection, reshape, end-to-end synthetic GGUF.

https://claude.ai/code/session_01Y69Vnw751w75iVSBRws7o7

Author: claude