feat: Add int8 KV cache compression with head-major layout and async pipelining

[naalo2]

Closes #228

Summary

Add int8 KV cache compression with head-major memory layout and async stream pipelining to hide KV store latency behind attention computation.

Key Changes

• config.py: Add quant flag and group_num configuration fields
• context.py / sequence.py: Update parameter passing interfaces and methods
• attention.py: Add quant computation branch and quant_store operator
• model_runner.py: Add quant KV cache initialization branch and CUDA graph capture under quant path
• tools/quant_attn_kvhead_based.py: Implement concrete int8 quantization compute operators

Benefits

• Reduced KV cache memory footprint via int8 quantization
• Improved GPU memory access efficiency via head-major coalesced layout
• Better throughput via async pipelining overlapping KV transfers with computation