feat(int4): add nibble_pack/unpack for true 0.5 B/elem int4 storage

[lokic233]

Summary

Addresses #328. qwix int4 QArrays currently store at 1.0 bytes/element (JAX's jnp.int4 byte-pads). This PR adds nibble_pack() / nibble_unpack() utilities that store two 4-bit values per uint8 byte — achieving the true 0.5 B/elem (4× vs fp32, 2× vs int8) memory reduction that int4 is meant to provide.

What's included

• qwix/_src/core/nibble_pack.py: nibble_pack(q) [N,K]→[N,K/2] uint8 and nibble_unpack(packed) [N,K/2]→[N,K] int8 in [-8,7]
• tests/nibble_pack_test.py: roundtrip exactness, memory halving, boundary values

Measured

• Verified on NVIDIA GB200 + Google TPU v7x (jax 0.10.2)
• nibble_pack achieves exactly 0.5 B/elem (vs 1.0 for raw jnp.int4)
• Roundtrip is bit-exact across the full [-8,7] range
• Used in production for a fused W4A16 matmul kernel on TPU (0.83× bf16, rel 0.0026)

Note

We acknowledge the repo's policy on external contributions. This is offered as a concrete, tested implementation of the fix described in #328, ready to adapt/integrate however the maintainers prefer.

— via Navi on behalf of @lokic233

[jshin1394]

Thanks for submitting this PR and sharing your benchmarks, @lokic233! Storing int4 at true 0.5 B/elem is a huge win for memory footprint, and it's great to see your production results for the fused W4A16 TPU kernel.

1. Before we merge this as a standalone utility, we'd love to learn more about your follow-up plan for integrating this into Qwix. Do you have a follow-up plan to integrate nibble_pack / unpack natively into QArray / qarray.py or checkpoint_util.py? Is the goal for Issue int4 QArray stores at 1.0 bytes/element (no nibble-packing) — zero memory benefit #328 is for qwix.quantize() and QArray to automatically store int4 at 0.5 B/elem out of the box?

2. Minor comment regarding file location: we'd suggest placing these helpers in qwix/_src/utils/packing.py (as pack_int4 / unpack_int4) so _src/core stays focused on core tensor engine primitives and we have a clean home for future bit-packing formats.

Looking forward to hearing your thoughts and roadmap for this!