feat(devices): make the virtio-fs backend state limit configurable

[0xpolarzero]

Lets the embedder choose how much virtio-fs backend state each device may capture and restore. Today the backend state is capped at 4 MiB and the encoded states at 8 MiB, all hard-coded, so a guest that has looked up a few tens of thousands of host files cannot be checkpointed. microsandbox wants to expose this as a setting (superradcompany/microsandbox#1751).

What changed

• VmBuilder::fs_state_limit(bytes) sets the backend state budget for every virtio-fs device, passthrough and custom. The default is DEFAULT_MAX_FS_BACKEND_STATE_BYTES (4 MiB, as before).
• The enclosing limits are derived from the budget instead of fixed at 8 MiB:
  • the fs device state is the 22-byte header plus the budget;
  • a VirtioDeviceState for virtio-fs adds the existing 1 MiB envelope allowance on top.
• VirtioDeviceState::encode_with_fs_backend_limit / decode_with_fs_backend_limit take the budget. encode / decode keep their signatures and use the default.
• max_virtio_device_state_bytes(device_type, budget) returns the largest encoded state, so embedders can size their own admission checks the same way.

Non-fs devices are unchanged. At the default budget, every state that 0.1.40 could capture still round-trips. The generic VirtioDeviceState bound for virtio-fs tightens from 8 MiB to about 5 MiB, which the fs device already enforced on restore.

How it was tested

• New unit tests in devices and vmm cover:
  • a state above the default budget is refused by default and round-trips with a larger budget;
  • exact header + budget bounds;
  • the device setter reaching capture, validate and restore;
  • non-fs limits ignoring the budget.
• cargo test for msb_krun_devices (virtio::fs), msb_krun_vmm --features blk and msb_krun --features net,blk, plus clippy -D warnings on those crates, on macOS.
• End to end through microsandbox (feat(snapshot): configure filesystem state limits microsandbox#1751, built against this branch) on macOS: a sandbox that read 50,000 files from a bind mount fails a full snapshot at the default budget and succeeds at 64 MiB. Restore, fork, verify and export/import of that snapshot also work.
• Not run here: the full-feature workspace clippy (needs libclang and pkg-config deps), and Linux/Windows builds.

Confidence Score: 5/5

The PR appears safe to merge based on the changes since the previous review.

What we checked:

• VM codec and device limits differ: The VM codec and both kinds of filesystem device read the limit placed in VM resources by the builder.

_{Reviews (3) · Last reviewed commit: "feat(api): share device state limits bet..."}

[toksdotdev]

@0xpolarzero made a minor change to share the device state limits, and just a few function renames.

[toksdotdev]

lgtm!