Add graph-safe 1-bit PQ KV cache support

[zzz0906]

Summary

• add compact INT1 and product-quantized K/V storage with per-layer codebooks
• integrate mixed BF16/quantized allocation, flush, prefill, and decode paths
• add CUDA Graph-safe Triton attention kernels and focused allocator/kernel coverage
• exclude experiment logs, generated codebooks, result artifacts, and the obsolete ternary path

Test plan

• Ruff syntax/undefined-name checks (E9, F821)
• Compile all changed Python files
• Source snapshot H100 validation: 43 tests plus 2 parameterized subtests; Qwen3-8B CUDA Graph batch 1/2/4
• Re-run the GPU suite in CI or a CUDA environment (the PR host has no PyTorch/CUDA)