fix(hicache): accept no-layout logical pools on the v2 registration path

[Wine93]

SGLang v0.5.17's DSV4 hybrid stack (build_deepseek_v4_hicache_stack -> attach_storage_backend) registers the logical anchor "kv" pool through register_mem_host_pool_v2 at attach time, and its
DeepSeekV4PagedHostPool.get_page_buffer_meta() declares "no host layout" as a 2-tuple containing None (the older LogicalHostPool returned None). The v2 registration probe predates the logical-anchor fix (60c0522) and hard-raised "pool component probe returned no layout" on both wire forms; the RuntimeError propagated out of attach_storage_backend and killed the whole SGLang scheduler, forcing users to disable HiCache (observed with DFKV 2.1.1/2.10.0/2.15.1 + deepseek-v4-flash-0731).

• _meta_has_no_layout(): None or a 2-tuple with None members is an explicit no-layout declaration; malformed shapes (wrong arity, empty/mismatched components) still fail closed.
• register_mem_host_pool_v2: a no-layout pool registers marker-only (_pool_logical, replicated=True so MLA follower ranks skip writes) instead of raising; re-registering with a real layout clears the flag.
• _v2_io: logical pools write the one-byte anchor markers on put (the same "kv" sub-keys as the v1 anchor path, riding _put_flat's exist gate and retry) and no-op complete on get so the hybrid controller proceeds to load the real side pools; a late no-layout declaration at transfer time flips the pool to marker-only the same way.
• batch_set_v1/batch_get_v1: the logical-anchor branch now accepts the tuple-with-None wire form too.

Test: FakeTupleNoneAnchorPool + 3 new cases (v1 tuple-None contract writes markers, v2 registration accepts both wire forms with set/get keeping marker semantics, physical re-registration clears the logical flag); test_dfkv_hicache 76 passed, device-direct 23 passed/10 skipped.