perf(ivfpq): eliminate PQ8 LUT bounds checks

[jerry-024]

Summary

Speed up the transposed 8-bit IVF-PQ scan by exposing each 256-entry PQ lookup table as a fixed-size array, allowing the compiler to eliminate bounds checks in the hot loop. The sparse/dense crossover remains unchanged at divisor 8 (12.5%), and search results are unchanged.

Depends on #75. After #75 merges, the effective diff for this PR starts at 1357928.

Changes

• Add a fixed [f32; 256] fast path for transposed PQ8 distance accumulation.
• Keep the existing slice-based fallback for non-PQ8 tables.
• Keep TRANSPOSED_SPARSE_SCAN_DIVISOR = 8.
• Update the filter-scan benchmark to use the production query shape (nq=64, nprobe=64) and deterministic randomized filtering.
• Cover the configured sparse/dense boundary and both fixed-array scan paths.

Testing

• cargo fmt --all -- --check
• cargo test -p paimon-vindex-core --lib (450 passed)
• cargo clippy -p paimon-vindex-core --lib --benches -- -D warnings
• Single-thread benchmark on the same SSH dev host with divisor 8 and the PQ8 optimization:

Filter density	p50
6.25%	541.739 ms
12.5%	907.507 ms
18.75%	1058.931 ms
25%	1040.456 ms
50%	1071.382 ms
100%	1066.809 ms

Historical scalar results from the same host and shape were 1890.700 ms at 50% and 1849.500 ms at 100%, corresponding to directional improvements of 43.3% and 42.3%. These runs were not interleaved A/B measurements; production impact still requires a Spark run on the target AMD executor.

Notes

• No unsafe code, architecture-specific SIMD, index-format change, or user-facing configuration is added.
• This PR does not change recall, filtering semantics, or the number of PQ codes evaluated.