Verilator --savable: serialize the context by pointer; checkpoint a tap's counters too

Verilator 5.x: os << *contextp does not compile (no operator for VerilatedContext&); the overload takes the pointer: VerilatedSave os; os.open(f); os << contextp << *topp << my_u64; and os >> contextp >> *topp >> my_u64 on restore (basic uint64_t/uint32_t operators exist, so you can append your own harness state). Put all TB state (memories, LFSRs, reset counters, plusarg-read values) in the Verilog model so it is saved; $value$plusargs results come back from the save. A restored run can reopen a VCD (tfp->open after restore) to dump a window, and if a C++ capture tap derives chunk indexes from change counts, write those counts into the same save stream or the restored run's chunks won't align. With a 64 KiB memory, picorv32 saves were 67 KB and ~0.3 ms each; restore+trace windows matched the original run's stored streams byte for byte. Also: Verilator's levels arg to top->trace() did not limit depth for me; use tfp->dumpvars(level, "TOP.tb") before open().