[MDA2AV]

Fixes #98.

Every io_uring_enter looked up the ring's own descriptor and took a reference to it (fget), then dropped it again (fput), before doing any work. That happens once per loop pass, on the process's busiest syscall. IORING_REGISTER_RING_FDS (kernel 5.18) puts the ring in the calling thread's table and hands back a slot. An enter that passes that slot with IORING_ENTER_REGISTERED_RING reads the ring from the table instead. liburing does this for every ring it creates.

What changes:

• Registration: Ring.Create registers the descriptor as its last step, and SubmitAndWait passes the slot and the flag.
• Same thread: the slot lives in the creating thread's table. That restricts nothing here: only the thread that created a SINGLE_ISSUER | DEFER_TASKRUN ring can enter it at all.
• Fallback: a thread has 16 slots. A ring created after they are all taken enters by descriptor, as before.
• Dispose: frees the slot before closing the descriptor. The slot holds a reference to the ring, so a ring closed while still registered would stay alive, memlock charge included, until its thread exited.

Ring.Fd is unchanged, and the register calls keep using it.

Measured

Per enter, with a C probe on a ring set up like ioxide's, pinned to one P-core, 5M iterations, 3 alternating rounds each:

	by descriptor	by registered slot
empty GETEVENTS enter	71 ns	64 ns	−7 ns
NOP round trip (submit 1, wait 1)	90 ns	82 ns	−8 ns

The same holds with a second thread alive, as in any .NET process.

wrk on Tcp/Raw: flat, as it should be when one enter carries a batch of requests.

• 2 reactors on two P-cores, wrk -t4 -c64 on other cores.
• Every round runs each build back to back, in rotated order; a second worktree of main is the noise control.
• Figures are the median of the per-round ratios.

run	rounds	req/s vs main	CPU/req vs main	control, main vs main
1	10 × 10 s	−1.5%	+1.2%	−0.2% / +0.5%
2	12 × 10 s	+0.8%	−1.5%	+0.6% / −0.1%

The two runs disagree in sign, and single rounds varied by up to ±10% on this box today.

Tests

New, in RegisteredRingTests (Unit):

• ring: enters go through the registered slot, not the descriptor: a NOP round trip works, then the slot is emptied behind the ring's back and the next enter must fail with EBADF. An enter by descriptor would not notice.
• ring: disposing a ring frees its registered slot: after Dispose, an enter on the old slot must answer EBADF.
• ring: a ring created with every slot taken still works, through its descriptor: 17 rings on one thread. The first takes a slot; the 17th falls back to its descriptor and works.

Red-checked, one break per test:

	enters by slot	dispose frees the slot	fallback
enters by descriptor (slot registered, unused)	FAIL: got 1, not EBADF	PASS	PASS
Dispose leaves the slot registered	PASS	FAIL: the closed ring still answers	PASS
no fallback without a free slot	PASS	PASS	FAIL
this PR	PASS	PASS	PASS

Docs: the ring setup list in architecture and the Ring section of dev-core.

All suites pass, every reactor in them on a registered ring: Unit 105, E2E 322, Chaos 80, Http 89, Tls 169 (+14 skipped), File 5, Pg 7 (+11 skipped), Redis (6 skipped, 1 pending). The Tls skips are the kTLS cases (no tls module loaded here), and the Pg and Redis ones need servers this box doesn't run.

It merges cleanly with #386, which also changes Ring.Create.

API: Native gains IORING_ENTER_REGISTERED_RING, IORING_REGISTER_RING_FDS, IORING_UNREGISTER_RING_FDS and io_uring_rsrc_update. Nothing is removed or changed.