[CPU] add support for mamba causal conv1d for qwen3-next

[mingfeima]

Motivation

add support for mamba causal conv1d for qwen3-next

Modifications

add kernel files at sgl-kernel/csrc/cpu/mamba/conv.cpp, which implemented both causal_conv1d_fwd for prefill and causal_conv1d_update for decode.

• APIs align with existing CUDA counterpart
• support both batched input and variant length input
• CPU kernels will requires x to be contiguous on the second-to-last dimension to achieve optimal performance
• shift conv_states to be contiguous the second-to-last dimension to make sure vectorized load from memory
• implemented with avx512-bf16, applying amx would be pointless since width is 4 (which conresponds to K in tinygemm of amx, it requires to be 32x)
• weight prepacked to be vnni2 format to remove online prepacking overhead

Accuracy Tests

python /test/srt/cpu/test_causal_conv1d.py

Benchmarking and Profiling

compare the optimized C++ version against reference implementation with native torch (which will end up with oneDNN)

performance collected on 6th gen Xeon with 40 cores, with script.

### batch = 1, dim = 8192, seqlen = 1024
### causal_conv1d: oneDNN ref: 3.205 ms; opt: 0.239 ms

### batch = 128, dim = 8192, seqlen = 1024
### causal_conv1d: oneDNN ref: 406.583 ms; opt: 36.051 ms

NOTE: the main reason for low performance with torch native implementation is that PyTorch doesn't have channels last concept for 1d convolution. On PyTorch, 1d convolution will be mapped to 2d and therefore it will always be channels first. This will trigger:

• addtional copy (tranpose last 2 dimensions) to make the input contiguous
• input and weight will need to be reordered to onednn internal format to use vnni
• output will be reordered from internal format to plain format

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[mingfeima]

fix new lint error.