Support JLD2 for (de)serialization

[lassepe]

Currently, serializing a CUDA array with JLD2 will silently fail because it serializes a pointer that is no longer valid when deserializing. This PR adds an extension for JLD2 that handles this more gracefully by hooking into the JLD2.writeas mechanism.

[github-actions]

Your PR no longer requires formatting changes. Thank you for your contribution!

[lassepe]

Buildkite failure seems unrelated.

[github-actions]

CUDA.jl Benchmarks

Benchmark suite	Current: 915565a	Previous: 1004fd9	Ratio
latency/precompile	43090085487 ns	43347615330 ns	0.99
latency/ttfp	7157885101 ns	7144432561 ns	1.00
latency/import	3445460616 ns	3422583362 ns	1.01
integration/volumerhs	9608976 ns	9621539.5 ns	1.00
integration/byval/slices=1	147297.5 ns	147204 ns	1.00
integration/byval/slices=3	425946 ns	426126.5 ns	1.00
integration/byval/reference	145338 ns	145306 ns	1.00
integration/byval/slices=2	286695.5 ns	286683 ns	1.00
integration/cudadevrt	103611 ns	103645 ns	1.00
kernel/indexing	14400 ns	14297 ns	1.01
kernel/indexing_checked	15055 ns	14899 ns	1.01
kernel/occupancy	715.6597222222222 ns	742.9652777777778 ns	0.96
kernel/launch	2247.6666666666665 ns	2214.8888888888887 ns	1.01
kernel/rand	15456 ns	18344.5 ns	0.84
array/reverse/1d	20207 ns	19882 ns	1.02
array/reverse/2d	25271 ns	23850 ns	1.06
array/reverse/1d_inplace	11469 ns	10508 ns	1.09
array/reverse/2d_inplace	13076 ns	12080 ns	1.08
array/copy	21159 ns	21206 ns	1.00
array/iteration/findall/int	160004.5 ns	159710.5 ns	1.00
array/iteration/findall/bool	139875 ns	139788 ns	1.00
array/iteration/findfirst/int	164540.5 ns	163426.5 ns	1.01
array/iteration/findfirst/bool	165590 ns	164898.5 ns	1.00
array/iteration/scalar	73700 ns	73632 ns	1.00
array/iteration/logical	218991.5 ns	219076 ns	1.00
array/iteration/findmin/1d	47585 ns	47341 ns	1.01
array/iteration/findmin/2d	97654 ns	97310 ns	1.00
array/reductions/reduce/1d	37152 ns	36812 ns	1.01
array/reductions/reduce/2d	41874 ns	41111 ns	1.02
array/reductions/mapreduce/1d	35265 ns	35000.5 ns	1.01
array/reductions/mapreduce/2d	41376 ns	41081 ns	1.01
array/broadcast	21459 ns	21375 ns	1.00
array/copyto!/gpu_to_gpu	12714 ns	12932 ns	0.98
array/copyto!/cpu_to_gpu	218545 ns	217571 ns	1.00
array/copyto!/gpu_to_cpu	286037 ns	286553 ns	1.00
array/accumulate/1d	109547 ns	109198 ns	1.00
array/accumulate/2d	81383 ns	81190 ns	1.00
array/construct	1256.3 ns	1248.9 ns	1.01
array/random/randn/Float32	44210 ns	44976.5 ns	0.98
array/random/randn!/Float32	25252 ns	25430 ns	0.99
array/random/rand!/Int64	27332 ns	27147 ns	1.01
array/random/rand!/Float32	8805 ns	8919.333333333334 ns	0.99
array/random/rand/Int64	30361 ns	30142 ns	1.01
array/random/rand/Float32	13382 ns	13307 ns	1.01
array/permutedims/4d	61774 ns	61252 ns	1.01
array/permutedims/2d	55763 ns	55576 ns	1.00
array/permutedims/3d	56487 ns	56240 ns	1.00
array/sorting/1d	2765868 ns	2778491 ns	1.00
array/sorting/by	3370345 ns	3369601 ns	1.00
array/sorting/2d	1087106 ns	1087382 ns	1.00
cuda/synchronization/stream/auto	1049.7 ns	1018.7 ns	1.03
cuda/synchronization/stream/nonblocking	7255.6 ns	8223 ns	0.88
cuda/synchronization/stream/blocking	830.5591397849462 ns	800.9072164948453 ns	1.04
cuda/synchronization/context/auto	1172.5 ns	1159.8 ns	1.01
cuda/synchronization/context/nonblocking	7133.1 ns	7478.9 ns	0.95
cuda/synchronization/context/blocking	921.1160714285714 ns	897.4426229508197 ns	1.03

This comment was automatically generated by workflow using github-action-benchmark.