simjoin: exact weighted-cosine similarity join (L2AP, CPU+GPU) + Python bindings (v0.3.0)

[prostomarkeloff]

Adds simjoin — an exact all-pairs weighted-cosine similarity join over sparse non-negative vectors (every pair with cos ≥ t, no approximation). The principled exact replacement for shingle-candidate + verify near-duplicate detection (Type-3 code clones: functions × IDF-weighted canonical lines).

What's in it

• src/simjoin.rs — SOTA L2AP (inverted index + Cauchy–Schwarz prefix prune) + branchless sorted-merge dot + cache-packed prune state + rayon parallel probe. Asserted bit-identical to an O(n²) brute-force oracle on fuzzed corpora (the crate's "two implementations, one answer" gate). API: Corpus::{from_rows, from_token_docs}, cosine_join, cosine_join_with(Concurrency), CosineJoiner handle.
• src/simjoin_gpu.rs — Metal batch_cosine kernel. The verify is memory-bandwidth-bound; the Apple GPU clears the random-gather dots ~3× the CPU (53 vs 22 GB/s). Metal is f32-only, so f64 byte-parity is preserved via GPU-filter + CPU-exact-reverify (gpu+cpu), plus a pure-f32 fast path (gpu, ≤1 differing pair per millions). ~1.8–2× end-to-end on the real top-300 PyPI corpus (287k functions, 3.1M clone pairs).
• 3 backends behind the existing Concurrency enum (cpu / gpu+cpu / gpu), transparent CPU fallback.
• Python: difflib_fast.cosine_join(docs, t, concurrency, threads) + CosineJoiner — token docs → TF-IDF in Rust, auto-parallel (rayon, GIL released) like ratio_many.
• examples/{simjoin_bench, simjoin_pypi, simjoin_gpu_bench}; profiling cargo feature; README + benchmarks.md §6.

Verification

• cargo test (CPU parity vs brute force) + cargo test --features gpu (gpu_hybrid_matches_cpu: exact backends byte-identical) — green.
• clippy clean on default / profiling / gpu / python,gpu; cargo publish --dry-run clean.

Version bump 0.2.0 → 0.3.0.