Reuse managed CPU execution coordination

[Ryo-wtnb11]

Why

A persistent managed CPU context still rebuilt execution coordination on every entry: immutable CPU domains were cloned, active arbiter requests used allocating tree nodes, and execution-owner state was broadcast to every worker on entry and exit. Faer policy prepared outside the owned pool could also capture an unrelated ambient Rayon degree.

This made empty warm entry pay costs unrelated to the numerical operation and made provider planning depend on where it was constructed.

Closes #1388.

What changed

• Store immutable CpuSet contents in shared Arc<[CpuId]> storage.
• Retain active-request capacity in the resource arbiter while preserving the fair waiter queue, conflict rules, reentry, and RAII release behavior.
• Register a shared worker execution scope once when a managed Rayon pool is constructed, then update its owner/depth without per-entry broadcasts.
• Derive Faer parallelism explicitly from the configured CpuContext degree, including policies created outside the managed pool.
• Add allocation, panic/reentry, worker-registration, provider-policy, source-contract, and benchmark coverage.
• Keep the TBLIS selection/fallback and provider-owned threading contract introduced by Add optional TBLIS CPU contraction provider #1344 unchanged.

Allocation boundary

The regression contract removes mandatory tenferro-owned allocation from compatible warm managed entry. It does not promise that Rayon never grows injector storage or that third-party providers never allocate. The sustained diagnostic observed occasional scheduler allocations after warmup, so the CI gate verifies that warm entry can complete without a mandatory allocation for one, two, and four configured workers rather than claiming unbounded global zero allocation.

Verification

• cpu-faer lib tests: 288 passed on macOS, including the portable cross-pool reentry fixture.
• Release install_allocation_tests: 1 passed.
• cargo check --no-default-features --features cpu-blas: passed.
• cargo check --no-default-features --features cpu-faer,cpu-tblis: passed with the dynamic-loading TBLIS feature. This is a compile check; an external TBLIS link/runtime was not exercised.
• TBLIS provider source-contract tests: 5 passed in independent review.
• cargo fmt --all -- --check and git diff --check: passed.
• Independent review of the final range found no P0-P2 issues.

Platform note

An earlier broader macOS CPU run on the pre-#1344 base reported five placement/NUMA failures caused by unsupported managed-affinity environment assumptions, not by numerical or execution-scope results. #1344 made those platform requirements explicit and cfg-gated the affected fixtures. On the current base, the macOS cpu-faer lib suite passes 288/288; the newly portable cross-pool fixture runs on macOS rather than being hidden by the placement gate.

[Ryo-wtnb11]

Superseded by a same-repository PR because the repository-rules workflow intentionally rejects external-head PRs. The failure was the IS_EXTERNAL_PR=true policy gate, not a code or test failure. The replacement uses the identical df4e6d46 HEAD/tree.