fix(v4): move hot rows16 packing out of store mutex

[aniketshukla1]

Summary

• Move rows16 packing behind a per-slot preparation mutex instead of the store-wide mutex.
• Convert in place only while the requesting lease is exclusive; an existing row-major view keeps its layout until a later lookup can safely repack the slot.
• Publish the packed state under the store mutex, so consumers of the same slot wait while unrelated expert loads continue.
• Add a hot-store regression test that verifies deferred conversion and concurrent packed lookups; it is part of make check.

Validation

• make check
• make tests/test_deepseek_v4_hot_store && ./tests/test_deepseek_v4_hot_store

Resolves #900