Fix 4-bit quantization for weight matrices not divisible by blocksize

[Abdennacer-Badaoui]

The Triton 4-bit quantization kernels assume the total number of elements is evenly divisible by blocksize. When this doesn't hold, the last block is partially filled with uninitialized data, which corrupts the absmax scaling factor for that block.

This PR fixes the issue by padding the input tensor with zeros to the next multiple of blocksize before entering the kernel. The padding is purely internal to quantize_4bit ; the output tensor and dequantization path use the original shape, so callers are unaffected. Zero-padding has no impact on absmax accuracy since max(abs(...)) is unaffected by additional zeros.

A roundtrip test is added to verify quantize/dequantize works correctly with non-divisible shapes and produces no NaN or Inf values.

[github-actions]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[matthewdouglas]

LGTM, thanks!