Fix spurious "Nested loops over ThisThreadPool" error

[tapios]

jl_in_threaded_region is a process-global flag, but parallelize_over(f, ::ThisThreadPool) read it twice. If an unrelated task entered a threaded region between the two reads, the first read said "not in a region" and the second said "in a region", leading to an error without actual nesting.

This affects any code that broadcasts Fields from several concurrent tasks. The plane/inertial_gravity_wave example does that via ThreadsX.mapreduce, and fails intermittently (PR #2556, build 7117).

CPU Multithreading & Scope Fixes

• Scope Resolution: Loops resolve their scope once up front with resolved_scope, and thread rank comes from task-local storage (:climacore_pool_threads) rather than global state.
• Thread Pool Partitioning: Concurrent loops divide the CPU pool dynamically, each taking an equal share of threads instead of letting the first loop claim all threads. This is ~2.6x faster with two concurrent tasks on 8 threads.
• Explicit Scope & Task Spawning: Explicit ThisThreadPool() calls now resolve properly. Pool threads are spawned explicitly into :default with @spawn :default (protecting against execution inside :interactive tasks), and parallel worker errors are collected and thrown via CompositeException.

GPU Performance Optimizations (ClimaCoreCUDAExt)

• Device Attribute Caching: Cache invariant CUDA device attributes (MAX_THREADS_PER_BLOCK, WARP_SIZE, SM_COUNT, etc.) to eliminate repeated C driver calls (cuDeviceGetAttribute) on every launch. Reduces host-side broadcast launch latency by ~54x (from ~477 μs down to ~8.8 μs per launch on an A100).
• Occupancy Query Caching: Cache invariant kernel launch configurations (CUDA.launch_configuration) in an IdDict to eliminate repeated C driver occupancy calculations (cuOccupancyMaxPotentialBlockSize) on every broadcast and reduction launch. Saves an additional ~2 μs of host CPU launch overhead per kernel call (~4.2% faster reductions).
• Task-Local Reduction Buffers: Intermediate GPU reduction arrays in reduce_points are cached in task-local storage to avoid allocating device memory (similar) during repeated scalar reductions.
• Warp Alignment: Ensure thread block sizes for slice loops are rounded to warp multiples (32) to prevent warp divergence.

CI & Workflow Improvements

• Documentation Deploy Concurrency: Added a concurrency group (documentation-deploy) to .github/workflows/docs.yml to queue documentation preview deployments sequentially, preventing git --force-with-lease rejections (! [rejected] HEAD -> gh-pages (stale info)) when multiple CI jobs run concurrently.