Add CUDA Backend Support

[mgyoo86]

Summary

GPU memory pooling with zero-allocation design, matching the existing CPU pool semantics.

Features

Zero-Allocation GPU Pooling

• @with_pool :cuda macro with automatic checkpoint/rewind lifecycle
• Task-local pools with multi-device awareness (one pool per GPU per Task)
• N-way view cache eliminates both GPU and CPU allocation after warmup

Unified API

• Same acquire! / unsafe_acquire! interface as CPU
• Returns CuArray{T,N} (not SubArray/ReshapedArray like CPU)
• Backend dispatch via Val{:backend} pattern for extensibility

Usage

using AdaptiveArrayPools, CUDA

@with_pool :cuda pool begin
    A = acquire!(pool, Float64, 100, 100)  # CuArray{Float64,2}
    B = acquire!(pool, Float64, 100, 100)
    fill!(A, 1.0); fill!(B, 2.0)
    sum(A .+ B)
end  # automatic rewind - arrays returned to pool

# Zero GPU allocation in hot loops
for i in 1:1000
    @with_pool :cuda p begin
        tmp = acquire!(p, Float32, 1000)
        # ... compute ...
    end
end

Implementation Details

Component	Description
CuTypedPool{T}	Type-specific pool with N-way view cache
CuAdaptiveArrayPool	Multi-type pool with 8 fixed slots + fallback
get_task_local_cuda_pool()	Per-Task, per-device pool retrieval
N-way cache	4-way set-associative cache per slot (configurable via CACHE_WAYS)

Allocation Behavior

• GPU: Always 0 bytes after warmup (backing CuVector resized and reused)
• CPU: 0 bytes on cache hit (≤4 dimension patterns), ~80 bytes on miss

[Copilot]

Pull request overview

This PR adds comprehensive CUDA backend support to AdaptiveArrayPools, enabling zero-allocation GPU memory pooling with the same API as the CPU implementation. The implementation follows a well-structured extension pattern and maintains backward compatibility while introducing GPU capabilities.

Key Changes

• Introduces abstract type hierarchy (AbstractTypedPool, AbstractArrayPool) to enable backend extensibility
• Adds backend dispatch via Val{:backend} pattern for compile-time resolution with @with_pool :cuda syntax
• Implements CUDA extension with unified N-way cache for all dimensions (returns CuArray instead of view types)
• Updates terminology from "arrays/vectors" to "slots" throughout for consistency

Reviewed changes

Copilot reviewed 33 out of 34 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
src/types.jl	Adds abstract type hierarchy for extensibility
src/acquire.jl	Refactors allocation to use dispatch points (allocate_vector, wrap_array)
src/macros.jl	Implements _get_pool_for_backend dispatch and backend-specific code generation
src/task_local_pool.jl	Adds CUDA pool stub functions
src/utils.jl	Updates display logic to support both CPU and CUDA pools, changes "arrays" to "slots"
ext/AdaptiveArrayPoolsCUDAExt/	Complete CUDA extension with types, dispatch, acquire logic, state management, and utilities
test/cuda/	Comprehensive test suite for CUDA functionality (extension, allocation, N-way cache, display)
test/test_backend_macro_expansion.jl	Tests for backend-specific macro expansion
docs/cuda.md, docs/safety.md	New documentation for CUDA backend and safety guide
README.md	Updated to highlight CUDA support and simplified introduction

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[codecov]

Codecov Report

❌ Patch coverage is 84.74576% with 18 lines in your changes missing coverage. Please review.
✅ Project coverage is 94.48%. Comparing base (a1bc284) to head (79299ac).
⚠️ Report is 23 commits behind head on master.

Files with missing lines	Patch %	Lines
src/macros.jl	81.13%	10 Missing ⚠️
src/utils.jl	81.81%	6 Missing ⚠️
src/types.jl	0.00%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master       #9      +/-   ##
==========================================
- Coverage   96.68%   94.48%   -2.20%     
==========================================
  Files           7        7              
  Lines         603      671      +68     
==========================================
+ Hits          583      634      +51     
- Misses         20       37      +17     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.