rt: shard the multi-thread inject queue to reduce remote spawn contention

[alex]

The multi-threaded scheduler's inject queue was protected by a single global mutex (shared with idle coordination state). Every remote task spawn — any spawn from outside a worker thread — acquired this lock, serializing concurrent spawners and limiting throughput.

This change introduces inject::Sharded, which splits the inject queue into up to 8 independent shards, each an existing Shared/Synced pair with its own mutex and cache-line padding.

Design:

• Push: each thread is assigned a home shard on first push (via a global counter) and sticks with it. This keeps consecutive pushes from one thread cache-local while spreading distinct threads across distinct locks.
• Pop: workers rotate through shards starting at their own index, skipping empty shards via per-shard atomic length. pop_n drains from one shard at a time to keep critical sections bounded.
• Shard count: capped at 8 (and 1 under loom). Contention drops off steeply past a handful of shards, and is_empty()/len() scan all shards in the worker hot loop.
• is_closed: a single Release atomic set after all shards are closed, so the shutdown check stays lock-free.

Random shard selection via context::thread_rng_n (as used in #7757 for the blocking pool) was measured and found to be 20-33% slower on remote_spawn at 8+ threads. The inject workload is a tight loop of trivial pushes where producer-side cache locality dominates: with RNG, a hot thread bounces between shard cache lines on every push; with sticky assignment it stays hot on one mutex and list tail. RNG did win slightly (5-9%) on single-producer benchmarks where spreading tasks lets workers pop in parallel, but not enough to offset the regression at scale.

The inject state is removed from the global Synced mutex, which now only guards idle coordination. This also helps the single-threaded path since remote pushes no longer contend with worker park/unpark.

Results on remote_spawn benchmark (12,800 no-op tasks, N spawner threads, 64-core box):

threads before after improvement
1 9.38 ms 7.33 ms -22%
2 14.94 ms 6.64 ms -56%
4 23.69 ms 5.34 ms -77%
8 34.81 ms 4.69 ms -87%
16 32.33 ms 4.54 ms -86%
32 30.37 ms 4.73 ms -84%
64 26.59 ms 5.34 ms -80%

rt_multi_threaded benchmarks: spawn_many_local -8%, spawn_many_remote_idle -7%, yield_many -1%, rest neutral.

Developed in conjunction with Claude.

[ADD-SP]

Is the inject queue still a FIFO queue?

[alex]

Approximately -- each of the queue shards is FIFO, but nothing attempts to ensure ordering cross-shard.