[KV Cache] Zero compressed attention storage pages

[Leonccaa]

Summary

• derive the zeroer's physical attention page layout from storage_block_size
• keep the existing virtual-block split for ordinary full attention pages
• cover ordinary and compressed layouts with real cache specs and a CUDA page-zeroing test

Problem

Compressed attention exposes a logical scheduler block that is larger than its
physical cache page. For example, Qwen3.8 Flash Next uses a logical block size
of 784 and a compression ratio of 8, so its compressed QSA cache stores 98 rows
per physical page.

KVBlockZeroer previously formed its virtual-page ratio from the logical block
size and the group's kernel block size. For a compressed cache this multiplied
the physical span by the compression ratio again, so a reused block could clear
the wrong physical range instead of exactly one compressed storage page.

Fix

Resolve the physical kernel block size and virtual-page ratio from the cache
spec before deriving segment spans:

• ordinary attention: keep the group kernel block size and its existing ratio
• compressed attention: use storage_block_size as one physical page with ratio 1
• reject layouts whose storage block is not divisible by the selected kernel block

Validation

• 3 passed, 1 skipped for the focused controller test selection (CUDA test skipped there)
• Ruff check and format check pass for both changed files
• CUDA smoke on a Tesla V100-PCIE-32GB:
  • ordinary layout: (16, 49)
  • compressed layout: (98, 1)
  • zeroing block 1 clears only physical page 1
• Qwen3.8 Flash Next 125B NVFP4, TP4 on 4x V100 completed model load,
  4 GiB/GPU KV-cache initialization, PIECEWISE+FULL CUDA Graph capture, and
  1K/4K generation with the equivalent patch before rebasing onto current main

[yangzhuxinyzx]

审计通过：压缩注意力零页逻辑按物理 storage_block_size 解析，普通页行为保持不变；锁定提交在最新 main 的隔离 worktree 上通过 13 项定向测试（含 V100 CUDA 实页清零），完整 pre-commit 与 DCO 通过。按锁定 SHA 合并。