allow 256-block quantization for tensors whose width isn't a multiple of 256

[lucaspirola]

llama-quantize silently falls back whenever a tensor's leading dim isn't a multiple of the block size, which means all the block-256 quants (the K/IQ family) just don't apply to those tensors. There's an existing issue about a --no-fallback flag (#26616), but I think teh underlying ask is a bit bigger than that.

Nemotron-3-Nano-30B-A3B experts are 1856/2688/3712 wide — none divisable by 256. The result is that every published sub-4-bit GGUF of these models is stuck at ~16.5–18 GB no matter what quant you ask for. Even IQ1_M, which should be tiny, lands around ~17.9 GB, because the expert tensors end up as ~4.5–5.2 bpw block-32 types instead. On a 16 GB card that's literally the diffrence between serving 262k+ context and not fitting at all.

draft PR 708-145#33 ("Allow 256-block quantization on unaligned tensors") pads the last partial block inside the ggml-quants kernels, drops the nx % qk_k fallback in llama-quant, and pads the activation in the graph (build_mul_mat_padded). It's general — any uneven tensor quantizes to its block ceiling (1856 → 8×256 blocks), the padded tail is zeros (d=0), so it's exact for any model