v0.1.0

First release of the Kimi K3 inference engine in portable C99.

Runs the full 93-layer model, 69 KDA and 24 Gated MLA layers, 896 routed
experts with top-16 selection, SiTU-GLU and native MXFP4 expert weights,
on a single CPU with no BLAS, no framework and no GPU.

Trunk streaming makes the memory budget a dial rather than a floor: the
same checkpoint runs in 8.24 GB and in 224 GB and produces byte-identical
output at every budget measured in between.

Verified without model weights. The checkpoint is 1.56 TB, so correctness
has to be checkable without it or it will not be checked.