Problem

When a dstack VM crashes (kernel panic with panic=1) or gets restarted, QEMU truncates serial.log on the next boot — losing all previous boot logs. This is especially painful in TDX VMs where non-resettable CPUs cause cpus are not resettable, terminating → rapid restart loops, making it impossible to diagnose the original panic.

Changes

vmm: preserve serial logs across restarts

• Before each VM restart, append serial.log content to serial.history.log with a timestamped ===== boot @ <rfc3339> ===== separator
• Cap serial.history.log at a configurable max size (default 4MB, trimmed from the front to keep most-recent data)
• Add the same boot separator timestamps to stdout.log and stderr.log so individual QEMU sessions are clearly delimited
• New config option serial_history_max_bytes (supports human-readable sizes: "4M", "512K")

ci: fix Docker Build Check on PRs

• github.sha for pull_request events is a temporary merge commit that only exists in GitHub's internal refs and is never pushed to the repo — Dockerfiles doing git clone && git checkout ${DSTACK_REV} fail with exit 128
• Use github.event.pull_request.head.sha || github.sha so PRs use the actual branch HEAD while push-to-branch events continue to use github.sha

Test plan

• Restart a VM, verify serial.history.log is created with previous boot's serial output and a ===== boot @ <timestamp> ===== separator
• Verify stdout.log and stderr.log contain boot separators between restarts
• Verify history truncation works when exceeding configured serial_history_max_bytes — tested with 1K limit, 88KB history correctly trimmed to 991 bytes keeping tail content
• Verify serial.log still works normally for the current boot