snapshot: re-anchor the mc timebase on start

[mach-kernel]

Changes

• Re-anchor last_host_ticks in MemoryController::start(), after the running.swap guard so a redundant start does not discard live elapsed time.
• Not in load_state: machine.rs:1448 runs it well before the bulk memory restore at 1503-1526, and restart_peripherals() is the last statement of load_snapshot_inner.
• cpu_cycle_acc and rpss_cycle_acc are left alone; power_on already zeroes them and runs first on every restore path.

Coverage / needs eyes

• mc::tests::start_reanchors_timebase stales the anchor by 100 s, calls start(), reads REG_RPSS_CTR. Reverting the one-line fix fails it at 1500000954 counts, the predicted 1.5e9 for 100 s at 50 MHz. With the fix the same read lands at 777-831 counts over four runs.
• cargo test --release --features lightning,rex-jit,chd: 358 passed, 11 ignored.
• Found via one update_timers step of cpu_cycles=16630691385 after a snapshot restore, RPSS_CTR 0x24d1b19 to 0xc88ead47.
• Does not cover a CPU-only pause. gdb_stub.rs:726/810 and the monitor cpu stop command stop the CPU without stopping the MC, so start() never runs and a held breakpoint is still billed on the next MC read. Worth a follow-up, left out to keep this to one thing.