Intern LowCardinality Row values to reduce string allocations

[Onyx2406]

Summary

Optimize LowCardinality(String) scanning by caching Row() results per dictionary index, reducing string allocations from O(N rows) to O(K unique values).

Fixes #1762

The problem

For LowCardinality(String) columns, Row() creates a new Go string for every row via byte-to-string conversion, even when many rows share the same dictionary value. For a column with 1M rows and 100 unique values, this means 1M string allocations instead of 100.

Benchmark from @bobrik in the issue:

• Baseline: 7406ms, 9282 MiB RAM (1M allocations)
• Optimized: 429ms, 1341 MiB RAM (K allocations)
• Speedup: 17x, Memory: 7x reduction

Fix (20 insertions, 1 file)

Add a rowCache map[int]any to LowCardinality that caches index.Row(idx) results by dictionary index. On the first access for a given index, the value is created and cached. Subsequent rows with the same index reuse the cached value.

• Cache only used for non-pointer access (ptr == false) — the common path
• Cache cleared on Reset() (between blocks)
• Lazy initialization — no allocation until first Row() call

Why this works

LowCardinality columns store data as:

• Dictionary (index): K unique values
• Keys: N row-to-dictionary mappings

Without caching: Row(i) → indexRowNum(i) → index.Row(idx) → new string every time
With caching: Row(i) → indexRowNum(i) → cache lookup → return cached string

Test plan

• go build ./lib/column/... passes
• CI (full test suite)
• Existing lowcardinality_test.go tests cover scanning behavior