criu: parallelize independent root siblings during TClone restore

[yiying-zhang]

Summary

• add an opt-in CRIU_TFORK_PARALLEL_SIBLINGS=N restore path;
• distribute eligible root children across a bounded temporary helper pool;
• use clone3(set_tid) plus syscall-only CLONE_PARENT to preserve requested
  PIDs and the restored parent graph;
• retain the serial path for small trees and unsupported clone semantics;
• make 0 an explicit off value for exact-binary ablation.

Why

At 100 processes, profiling measured about 166.96 ms per run in
fork_with_pid(), including 155.75 ms in clone3. Kernel sampling assigned
most sampled copy_process() time to dup_mmap(). CRIU currently launches
root siblings serially even though their Linux task creation is independent.

This change overlaps that real kernel and restore initialization work instead
of optimizing another small image-decoding or procfs leaf.

Safety scope

The fast path is disabled by default and requires:

• active TClone restore at the process-tree root;
• clone3(set_tid) support;
• at least 32 children in the relevant session group;
• no child with CLONE_PARENT, CLONE_NEWPID, CLONE_VM, or
  CLONE_THREAD requirements.

Worker count is capped at 16. Nested restore tasks stay serial, so wide nested
trees cannot recursively multiply helper pools. The restore root blocks
SIGCHLD, reaps every temporary helper, and fails restore if any helper fails.
An ineligible group falls back before any parallel work begins.

Performance

The A/B used the exact same Podman, crun, libcriu, and CRIU files. Only
CRIU_TFORK_PARALLEL_SIBLINGS=0 versus 8 changed. Every run used a fresh
source, alternating AB/BA order, and a one-second idle cooldown.

Processes	Serial median	Parallel median	Paired median saving	Wins
1	506.682 ms	506.380 ms	-21.938 ms	1/5
4	546.294 ms	546.700 ms	7.864 ms	3/5
25	678.479 ms	684.028 ms	-10.993 ms	2/5
50	822.954 ms	770.983 ms	80.073 ms	5/5
100	1,127.787 ms	1,030.824 ms	90.810 ms	10/10
200	1,762.010 ms	1,547.472 ms	236.945 ms	5/5

At 100 processes, the internal Phase-B-marker-to-clone-tree median fell from
343.010 to 227.406 ms, with 118.879 ms paired-median saving. A median fit
over all six points changes from 6.231 to 5.117 ms/process for this synthetic
workload.

Validation

• full CRIU build succeeds;
• criu check --all reports Looks good.;
• all final 50/100/200 process candidate samples won;
• every recorded source and child had the exact expected process count and
  flat parent graph;
• sentinel and source/child filesystem isolation checks passed;
• a 161-process nested tree preserved 40 direct parents and 120 grandchildren;
• 40 independent session leaders preserved their PID, PPID, and SID
  relationships;
• the source remained running and unchanged after every recorded fork.

Known limitations

The branch is intentionally opt-in pending nested PID-namespace,
shared-VM/thread-group, helper-failure-injection, concurrency, and long-soak
coverage. Three exploratory runs also hit the pre-existing 60-second
event-readiness timeout—two serial and one parallel. Transactional rollback
cleaned each unpublished child. This branch does not attempt to hide or retry
that separate reliability issue.