sandbox: allow restrictive nested seccomp filters without BPF JIT

[zbarsky-openai]

Sandboxed programs that install their own restrictive seccomp policy currently fail
inside actiond even though the worker already sets no_new_privs:

NoNewPrivs: 1
Seccomp: 2
prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER, ...): EPERM

Permit both supported nested-filter entry points (seccomp(SECCOMP_SET_MODE_FILTER)
and prctl(PR_SET_SECCOMP, SECCOMP_MODE_FILTER)). Disable the ARM64 kernel's BPF
JIT first so these additional restrictive filters cannot expose a classic-BPF JIT
surface; the x86_64 kernel already has its JIT disabled. Keep the existing
denials for bpf(), io_uring_setup, AF_VSOCK, and socket BPF attachment.

Add focused policy coverage and an end-to-end action-tool assertion that installs
an additional filter after verifying the original restrictions.

Validation:

• Resolved both ARM64 and x86_64 Linux configurations with the actual kernel
  Kconfig parser: CONFIG_SECCOMP_FILTER=y, CONFIG_BPF_JIT unset.
• Rebuilt the ARM64 kernel through actiond: 3,601 actions, 962 actually executed
  remotely (3d5ca79e-06b2-4409-ae92-cfcefef3551f).
• Rebuilt the ARM64 guest initramfs and booted the real macOS-hosted VM without
  adding packages or replacing its existing CAS.
• A freshly executed guest probe changed nested-filter installation from EPERM
  to success (ba4a5aae-c440-4247-9fa1-3f802dba9b23).
• Five seccomp policy checks and five sandbox-tool checks passed inside the new
  VM, including the retained BPF, socket, io_uring, and audit-architecture
  restrictions (47cc7191-684d-4c5d-b17a-31ee7c44088c).
• //src:unit_tests and //tools:e2e_action_tool_tests both executed and passed
  with test caching disabled (ae8fe693-bbc5-4c24-9d38-5a425f5358d0).