[3.3] GPU dictionary indexing — the reason this wave exists

[CSSFrancis]

Normalized cross-correlation between P experimental and D dictionary patterns is, once both are zero-mean / unit-norm, a single matmul E @ Dᵀ → (P, D), followed by top-k.

• Chunk over both P and D with a running top-k so the (P, D) intermediate never materialises — 65k × 100k float32 is 26 GB whole.
• kikuchipy does this with dask; torch should be dramatically faster.
• Support NormalizedCrossCorrelationMetric and NormalizedDotProductMetric semantics.

Acceptance: score and index parity against kikuchipy.indexing on the same dictionary, plus a benchmark at real scan scale.

Part of #68.

---

Plan: RELEASE_0_3_0_PLAN.md

[CSSFrancis]

Delivered in 0.3.0, merged to main via #89.

1216596 — feat(ebsd): GPU dictionary indexing + pattern preprocessing

Closing from the merge rather than automatically: the commits cite issues as (#71) rather than Closes #71, so GitHub linked none of them (#89 reported 0 closing references).