Drop DLPack dependency; reorganize examples as perf benchmarks

[CarloLucibello]

Summary

Two independent, docs-heavy changes bundled together:

1. Drop the DLPack.jl dependency

numpy2jl no longer routes through DLPack. Instead it wraps the numpy buffer as a genuine Array via a vendored PyArray + unsafe_wrap, rooted through a module-level WeakKeyDict (weak key = the wrapper Array, strong value = the PyArray) so the Python buffer stays alive exactly as long as the Julia array and is released when it's GC'd.

• Same behavior & performance as DLPack: the result is <: DenseArray, so BLAS/GPU/fast-indexing paths are preserved (a bare PyArray misses all of these). Axis reversal and the read-only / non-contiguous copy guards are kept.
• Bonus deadlock fix: PythonCall's GIL-deferred decref means a finalizer firing on a parallel=true worker thread enqueues the pointer rather than re-acquiring the GIL — so the thread-finalizer deadlock that DLPack's eager-decref finalizer could cause is gone.
• Removes the unmaintained DLPack dep from Project.toml and the using DLPack import.

2. Reorganize the Flux examples as perf/ benchmarks

• Move examples/flux_mnist and examples/flux_cifar10 under perf/ (alongside perf/load_dataset/) and delete the now-empty examples/. All doc/README path references updated.
• Add a 2-worker Distributed row to every benchmark config, run the Julia loaders at -t4 (was -t8), and make the PyTorch scripts read EPOCHS from the env.
• Refresh both READMEs with fresh 10-epoch numbers (Ryzen Threadripper PRO 9955WX CPU / RTX 5090 GPU). The longer runs give clean, monotonic worker-scaling curves.
• Fix a stale map(copy, …) note in the MNIST README (the code uses [:]) and align the deadlock framing with the deferred-decref behavior.

Testing

• Full test suite green (~447 tests), including the 15 serialization tests.
• Verified no memory leak (WeakKeyDict drains to 0) and no deadlock under parallel finalizers.
• Ran all four benchmarks end-to-end at 10 epochs (exit 0): MNIST ~98% test acc, CIFAR-10 ~80–83%.

🤖 Generated with Claude Code