Fix TCP RX softirq wake deadlock

[ryanbreen]

Summary

• Mask TCP global connection/listener/ISN locks with irq_save/irq_restore so same-CPU NetRx softirq cannot reenter while a thread owns a TCP lock.
• Keep NetRx wakeups lock-free by using the scheduler ISR wake buffer from TCP softirq wake paths.
• Extend the ISR wake-buffer drain to also wake legacy ThreadState::Blocked socket waiters, preserving existing syscall blocking semantics without editing syscall/*.

Root Cause

The RX stall was a same-CPU lock reentry deadlock. A userspace TCP syscall path could hold TCP_CONNECTIONS or TCP_LISTENERS, then a NIC IRQ/NetRx softirq ran on the same CPU and entered handle_tcp(), which tried to take the same TCP lock. The interrupted owner could not resume, so NetRx stayed in-progress with NET_RX_PROCESSING_HELD=1, NET_RX_SOFTIRQ_ENTRY_TOTAL = EXIT_TOTAL + 1, and NET_PCI_RX_USED_IDX ahead of NET_PCI_RX_LAST_USED_IDX.

Negative control artifacts:
/Users/wrb/Downloads/Ralph/breenix-interrupt-io-roadmap-1780056222/turn94-artifacts/negative-control/

Proof

Build gate:

• cargo build --release --features testing,external_test_bins --bin qemu-uefi passed clean with zero warning/error lines.

Required Parallels proof, same boot, no VM restart:

• Artifacts: /Users/wrb/Downloads/Ralph/breenix-interrupt-io-roadmap-1780056222/turn94-artifacts/fixed-proof-3/
• Ran 4 back-to-back cycles of outbound bssh 10.0.1.210 22 wrb --publickey --exec uname followed immediately by inbound host -> Breenix ssh ... 'echo AFTER && uname'.
• All 4 outbound runs returned real Darwin, rc=0.
• All 4 inbound runs returned AFTER and Breenix uname, rc=0.

Counter evidence from the fixed run:

• sample 3: NET_RX_SOFTIRQ_ENTRY_TOTAL=110, NET_RX_SOFTIRQ_EXIT_TOTAL=110, NET_RX_PROCESSING_HELD=0, NET_PCI_RX_USED_IDX=180, NET_PCI_RX_LAST_USED_IDX=180.
• sample 4: NET_RX_SOFTIRQ_ENTRY_TOTAL=237, NET_RX_SOFTIRQ_EXIT_TOTAL=237, NET_RX_PROCESSING_HELD=0, NET_PCI_RX_USED_IDX=374, NET_PCI_RX_LAST_USED_IDX=374.

Additional check:

• cargo run -p xtask -- boot-stages was attempted after the proof. It passed 63 stages including the softirq stages, then timed out on the existing x86 ARP gateway marker ([22/252] ARP reply received and gateway MAC resolved). QEMU was killed afterward.