machine: snapshot master FPU at prepare time for cross-process forks

[perbu]

Problem

Fork construction (Machine(const Machine&, const MachineOptions&)) and fork
reset (reset_to) copy the master's FPU state with a live KVM_GET_FPU against
the master's vCPU fd. That only works in the process that created the VM: KVM
vCPU ioctls require the caller's mm to be the VM-creating mm. A child that
inherited the master over fork() and builds its own VM from the master's
copy-on-write memory gets -EIO on that ioctl and cannot construct a fork.

Change

Snapshot the master's FPU registers into the Machine at
prepare_copy_on_write() time — while the master's vCPU fd is still ioctl-able —
and have fork construction and reset_to() read that snapshot via the new
prepared_fpu_registers() accessor instead of a live KVM_GET_FPU.

• Behavior-preserving for the in-process case. The master is frozen after
  prepare and never runs again, so the snapshot is bit-identical to what a live
  KVM_GET_FPU would return at fork time.
• Precondition made explicit. prepared_fpu_registers() asserts m_prepped,
  and the snapshot is only taken in prepare_copy_on_write().
• sregs guard. The special-registers half of cross-process safety is already
  covered by TINYKVM_USE_SYNCED_SREGS (reads the mmap'd kvm_run page instead
  of KVM_GET_SREGS). Added a #error guard so disabling that define can't
  silently reintroduce the -EIO on the sregs path.

Cost is one tinykvm_fpuregs (~fxsave/xsave layout) per Machine, always.

Testing

• cmake --build build clean (only pre-existing sign-compare warnings in
  src/tests.cpp).
• tests/run_unit_tests.sh: test_reset passes 46/47 assertions. The one
  failing assertion (output_is_hello_world) also fails identically on master
  — it comes from a test that compiles a static guest ELF with gcc at runtime
  and is environmental on my host, unrelated to this change. Same for
  test_tegridy. All other suites pass.

Motivation

Enables a process-per-VM model where a single-threaded zygote fork()s worker
processes that each construct a TinyKVM fork from a master inherited over
fork() — isolating tenant workloads in separate address spaces / jailed
processes rather than threads.

[fwsGonzo]

Does this break ARM? I thought I saw that ARM didn't support this?

[perbu]

yes. this is true.

checked on native ARM64. The existing ARM test suite passed, but I found one ARM-specific gap: PR 85 only snapshotted prepared FPU state in the non-ARM vcpu.cpp; ARM’s prepare_copy_on_write() was still leaving the prepared FP/SIMD snapshot zeroed. I pushed bb38a40, which snapshots it on ARM too and adds a regression test for fork construction and reset_to(). /tests/run_unit_tests.sh passes on ARM64.

sloppy.