Reduce redundant Top-K host setup

[morluto]

Fixes #7.

Reduce repeated CPU setup before Top-K launches without changing device-side computation or caching tensor descriptors.

Cache the tensor-map encoder pointer using thread-safe local-static initialization, retrying lookup if initialization throws. Add a fixed-rank metadata path while preserving the vector helper interface. Route Top-K's rank-three metadata through stack arrays, and remove duplicate attribute setters from the three launchers; the common launcher still sets and checks the attributes for each invocation.

The host regression harness uses actual CUDA types and product headers with a recording mock resolver/encoder. It checks concurrent cold initialization, sharing across two translation units, lookup failure retries, encoder failure recovery, fresh per-call arguments, and vector/array equivalence for ranks 1–5. The same test against baseline headers fails the intended single-lookup assertion.

With 200,000 calls per round, the baseline metadata path makes 1,600,000 allocations and 200,000 lookups; the candidate makes zero of either after warmup. Across 14 rounds per variant in baseline/candidate/candidate/baseline order, median instrumented host preparation time was 457 ns versus 114 ns on an Intel i7-10875H using NVCC 13.3, -O2, and libstdc++. Allocation accounting and a mock encoder are included in these timings: they are not driver-encoding or end-to-end Top-K results. Reproduction commands are in tests/host/README.md.

Validation also includes SM100a compilation of one specialization from each of the BF16, FP32, and BF16-cluster kernel families, plus whitespace checks.