fix sysvar handling when we’re slot leader & for speculative replay

[smcio]

We were using a global sysvar cache, which could refer to sysvar values from the wrong bank during speculative fork replay and when producing blocks as leader.

changes:

• introduced BankSysvars
• made replay and block production derive sysvars from the exact parent bank
• hence moved sysvar reads and other consumers away from the global cache
• alloc and deserialization free
• preserved sysvar state when passing the latest replayed bank state to block production and fork unwinds
• tests