feat: Implement 4-bit and 8-bit Quantized Convolutional Layers

[OnePunchMonk]

This PR introduces support for 4-bit and 8-bit quantized convolutional layers as requested in #1414.

Included:

• \Conv1d4bit, \Conv2d4bit, \Conv3d4bit\ with FP4 and NF4 types
• \Conv1d8bitLt, \Conv2d8bitLt, \Conv3d8bitLt\
• Complete test suite matching bitsandbytes standards
• Resolves the missing support for vision and audio models relying natively on convolutions.

Fixes #1414

[matthewdouglas]

Thanks for the contribution and the effort that went into this — the code is well-structured and the test coverage is thorough.

After reviewing, I don't think this is something we should add to bitsandbytes right now, and I want to explain why rather than just closing it.

Both the 4-bit and 8-bit conv classes here do dequantize-then-compute: weights are stored in reduced precision but fully dequantized back to fp16/fp32 before the forward pass, delegating to standard F.conv*d. For the 8-bit case in particular this creates a misleading class name — unlike Linear8bitLt which actually runs int8 tensor core arithmetic, Conv8bitLt doesn't do any int8 compute. The memory savings are real but achievable today with Params4bit and a custom forward.

The broader question is who this is for. The dominant use case for bitsandbytes is transformer models with large linear layers. The one place transformer architectures use something conv-shaped — GPT-2's Conv1D — is already handled by transposing the weight to make it compatible with the standard linear quantization path. The audience this PR targets is conv-heavy models like vision and audio backbones, which is a genuinely different audience with different memory profiles and tooling expectations. We haven't seen meaningful demand signal from that audience — the linked issue has a single reaction and no comments — and adding 12 new public classes is a long-term maintenance commitment we want to be confident about before taking on.

I'm going to close this PR for now. Thanks again for putting the work in.